usearch 2.21.4 → 2.25.1

package/numkong/include/numkong/mesh/haswell.h

@@ -0,0 +1,2013 @@
+/**
+ *  @brief SIMD-accelerated Point Cloud Alignment for Haswell.
+ *  @file include/numkong/mesh/haswell.h
+ *  @author Ash Vardanian
+ *  @date December 27, 2025
+ *
+ *  @sa include/numkong/mesh.h
+ *
+ *  @section haswell_mesh_instructions Key AVX2 Mesh Instructions
+ *
+ *      Intrinsic               Instruction                     Haswell         Genoa
+ *      _mm256_fmadd_ps         VFMADD (YMM, YMM, YMM)          5cy @ p01       4cy @ p01
+ *      _mm256_hadd_ps          VHADDPS (YMM, YMM, YMM)         7cy @ p1+p5     4cy @ p123+p23+p23
+ *      _mm256_permute2f128_ps  VPERM2F128 (YMM, YMM, YMM, I8)  3cy @ p5        2cy @ p12
+ *      _mm256_extractf128_ps   VEXTRACTF128 (XMM, YMM, I8)     3cy @ p5        1cy @ p0123
+ *      _mm256_i32gather_ps     VGATHERDPS (YMM, M, YMM, YMM)   22cy (34 uops)  19cy (17 uops)
+ *
+ *  Point cloud operations (centroid, covariance, Kabsch alignment) use gather instructions for
+ *  stride-3 xyz deinterleaving. Multiple FMA accumulators hide the 5-cycle FMA latency. VHADDPS
+ *  interleaves results across lanes, requiring additional shuffles for final scalar reduction.
+ */
+#ifndef NK_MESH_HASWELL_H
+#define NK_MESH_HASWELL_H
+
+#if NK_TARGET_X8664_
+#if NK_TARGET_HASWELL
+
+#include "numkong/types.h"
+#include "numkong/dot/haswell.h"
+#include "numkong/mesh/serial.h"
+#include "numkong/reduce/haswell.h"  // `nk_reduce_add_f32x8_haswell_`
+#include "numkong/spatial/haswell.h" // `nk_f32_sqrt_haswell`, `nk_f64_sqrt_haswell`
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+#if defined(__clang__)
+#pragma clang attribute push(__attribute__((target("avx2,f16c,fma,bmi,bmi2"))), apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC target("avx2", "f16c", "fma", "bmi", "bmi2")
+#endif
+
+/*  Deinterleave 24 floats (8 xyz triplets) into separate x, y, z vectors.
+ *  Uses AVX2 gather instructions for clean stride-3 access.
+ *
+ *  Input: 24 contiguous floats [x0,y0,z0, x1,y1,z1, ..., x7,y7,z7]
+ *  Output: x[8], y[8], z[8] vectors
+ */
+NK_INTERNAL void nk_deinterleave_f32x8_haswell_(nk_f32_t const *ptr, __m256 *x_out, __m256 *y_out, __m256 *z_out) {
+    // Gather indices: 0, 3, 6, 9, 12, 15, 18, 21 (stride 3)
+    __m256i idx_i32x8 = _mm256_setr_epi32(0, 3, 6, 9, 12, 15, 18, 21);
+    *x_out = _mm256_i32gather_ps(ptr + 0, idx_i32x8, 4);
+    *y_out = _mm256_i32gather_ps(ptr + 1, idx_i32x8, 4);
+    *z_out = _mm256_i32gather_ps(ptr + 2, idx_i32x8, 4);
+}
+
+/*  Deinterleave 12 f64 values (4 xyz triplets) into separate x, y, z vectors.
+ *  Uses scalar extraction for simplicity as AVX2 lacks efficient stride-3 gather for f64.
+ *
+ *  Input: 12 contiguous f64 [x0,y0,z0, x1,y1,z1, x2,y2,z2, x3,y3,z3]
+ *  Output: x[4], y[4], z[4] vectors
+ */
+NK_INTERNAL void nk_deinterleave_f64x4_haswell_(nk_f64_t const *ptr, __m256d *x_out, __m256d *y_out, __m256d *z_out) {
+    nk_f64_t x0 = ptr[0], x1 = ptr[3], x2 = ptr[6], x3 = ptr[9];
+    nk_f64_t y0 = ptr[1], y1 = ptr[4], y2 = ptr[7], y3 = ptr[10];
+    nk_f64_t z0 = ptr[2], z1 = ptr[5], z2 = ptr[8], z3 = ptr[11];
+
+    *x_out = _mm256_setr_pd(x0, x1, x2, x3);
+    *y_out = _mm256_setr_pd(y0, y1, y2, y3);
+    *z_out = _mm256_setr_pd(z0, z1, z2, z3);
+}
+
+/* Horizontal reduction helpers moved to reduce.h:
+ * - nk_reduce_add_f32x8_haswell_
+ * - nk_reduce_add_f64x4_haswell_
+ */
+
+NK_INTERNAL nk_f64_t nk_reduce_stable_f64x4_haswell_(__m256d values_f64x4) {
+    nk_b256_vec_t values;
+    values.ymm_pd = values_f64x4;
+    nk_f64_t sum = 0.0, compensation = 0.0;
+    nk_accumulate_sum_f64_(&sum, &compensation, values.f64s[0]);
+    nk_accumulate_sum_f64_(&sum, &compensation, values.f64s[1]);
+    nk_accumulate_sum_f64_(&sum, &compensation, values.f64s[2]);
+    nk_accumulate_sum_f64_(&sum, &compensation, values.f64s[3]);
+    return sum + compensation;
+}
+
+NK_INTERNAL void nk_rotation_from_svd_f64_haswell_(nk_f64_t const *svd_u, nk_f64_t const *svd_v, nk_f64_t *rotation) {
+    nk_rotation_from_svd_f64_serial_(svd_u, svd_v, rotation);
+}
+
+NK_INTERNAL void nk_accumulate_square_f64x4_haswell_(__m256d *sum_f64x4, __m256d *compensation_f64x4,
+                                                     __m256d values_f64x4) {
+    __m256d product_f64x4 = _mm256_mul_pd(values_f64x4, values_f64x4);
+    __m256d product_error_f64x4 = _mm256_fmsub_pd(values_f64x4, values_f64x4, product_f64x4);
+    __m256d tentative_sum_f64x4 = _mm256_add_pd(*sum_f64x4, product_f64x4);
+    __m256d virtual_addend_f64x4 = _mm256_sub_pd(tentative_sum_f64x4, *sum_f64x4);
+    __m256d sum_error_f64x4 = _mm256_add_pd(
+        _mm256_sub_pd(*sum_f64x4, _mm256_sub_pd(tentative_sum_f64x4, virtual_addend_f64x4)),
+        _mm256_sub_pd(product_f64x4, virtual_addend_f64x4));
+    *sum_f64x4 = tentative_sum_f64x4;
+    *compensation_f64x4 = _mm256_add_pd(*compensation_f64x4, _mm256_add_pd(sum_error_f64x4, product_error_f64x4));
+}
+
+/*  Compute sum of squared distances after applying rotation (and optional scale).
+ *  Used by kabsch (scale=1.0) and umeyama (scale=computed_scale).
+ *  Returns sum_squared, caller computes sqrt(sum_squared / n).
+ */
+NK_INTERNAL nk_f64_t nk_transformed_ssd_f32_haswell_(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n,
+                                                     nk_f64_t const *r, nk_f64_t scale, nk_f64_t centroid_a_x,
+                                                     nk_f64_t centroid_a_y, nk_f64_t centroid_a_z,
+                                                     nk_f64_t centroid_b_x, nk_f64_t centroid_b_y,
+                                                     nk_f64_t centroid_b_z) {
+    __m256d scaled_rotation_x_x_f64x4 = _mm256_set1_pd(scale * r[0]);
+    __m256d scaled_rotation_x_y_f64x4 = _mm256_set1_pd(scale * r[1]);
+    __m256d scaled_rotation_x_z_f64x4 = _mm256_set1_pd(scale * r[2]);
+    __m256d scaled_rotation_y_x_f64x4 = _mm256_set1_pd(scale * r[3]);
+    __m256d scaled_rotation_y_y_f64x4 = _mm256_set1_pd(scale * r[4]);
+    __m256d scaled_rotation_y_z_f64x4 = _mm256_set1_pd(scale * r[5]);
+    __m256d scaled_rotation_z_x_f64x4 = _mm256_set1_pd(scale * r[6]);
+    __m256d scaled_rotation_z_y_f64x4 = _mm256_set1_pd(scale * r[7]);
+    __m256d scaled_rotation_z_z_f64x4 = _mm256_set1_pd(scale * r[8]);
+    __m256d centroid_a_x_f64x4 = _mm256_set1_pd(centroid_a_x), centroid_a_y_f64x4 = _mm256_set1_pd(centroid_a_y);
+    __m256d centroid_a_z_f64x4 = _mm256_set1_pd(centroid_a_z), centroid_b_x_f64x4 = _mm256_set1_pd(centroid_b_x);
+    __m256d centroid_b_y_f64x4 = _mm256_set1_pd(centroid_b_y), centroid_b_z_f64x4 = _mm256_set1_pd(centroid_b_z);
+    __m256d sum_squared_f64x4 = _mm256_setzero_pd();
+    __m256 a_x_f32x8, a_y_f32x8, a_z_f32x8, b_x_f32x8, b_y_f32x8, b_z_f32x8;
+    nk_size_t index = 0;
+
+    for (; index + 8 <= n; index += 8) {
+        nk_deinterleave_f32x8_haswell_(a + index * 3, &a_x_f32x8, &a_y_f32x8, &a_z_f32x8),
+            nk_deinterleave_f32x8_haswell_(b + index * 3, &b_x_f32x8, &b_y_f32x8, &b_z_f32x8);
+
+        __m256d a_x_low_f64x4 = _mm256_cvtps_pd(_mm256_castps256_ps128(a_x_f32x8));
+        __m256d a_x_high_f64x4 = _mm256_cvtps_pd(_mm256_extractf128_ps(a_x_f32x8, 1));
+        __m256d a_y_low_f64x4 = _mm256_cvtps_pd(_mm256_castps256_ps128(a_y_f32x8));
+        __m256d a_y_high_f64x4 = _mm256_cvtps_pd(_mm256_extractf128_ps(a_y_f32x8, 1));
+        __m256d a_z_low_f64x4 = _mm256_cvtps_pd(_mm256_castps256_ps128(a_z_f32x8));
+        __m256d a_z_high_f64x4 = _mm256_cvtps_pd(_mm256_extractf128_ps(a_z_f32x8, 1));
+        __m256d b_x_low_f64x4 = _mm256_cvtps_pd(_mm256_castps256_ps128(b_x_f32x8));
+        __m256d b_x_high_f64x4 = _mm256_cvtps_pd(_mm256_extractf128_ps(b_x_f32x8, 1));
+        __m256d b_y_low_f64x4 = _mm256_cvtps_pd(_mm256_castps256_ps128(b_y_f32x8));
+        __m256d b_y_high_f64x4 = _mm256_cvtps_pd(_mm256_extractf128_ps(b_y_f32x8, 1));
+        __m256d b_z_low_f64x4 = _mm256_cvtps_pd(_mm256_castps256_ps128(b_z_f32x8));
+        __m256d b_z_high_f64x4 = _mm256_cvtps_pd(_mm256_extractf128_ps(b_z_f32x8, 1));
+
+        __m256d centered_a_x_low_f64x4 = _mm256_sub_pd(a_x_low_f64x4, centroid_a_x_f64x4);
+        __m256d centered_a_x_high_f64x4 = _mm256_sub_pd(a_x_high_f64x4, centroid_a_x_f64x4);
+        __m256d centered_a_y_low_f64x4 = _mm256_sub_pd(a_y_low_f64x4, centroid_a_y_f64x4);
+        __m256d centered_a_y_high_f64x4 = _mm256_sub_pd(a_y_high_f64x4, centroid_a_y_f64x4);
+        __m256d centered_a_z_low_f64x4 = _mm256_sub_pd(a_z_low_f64x4, centroid_a_z_f64x4);
+        __m256d centered_a_z_high_f64x4 = _mm256_sub_pd(a_z_high_f64x4, centroid_a_z_f64x4);
+        __m256d centered_b_x_low_f64x4 = _mm256_sub_pd(b_x_low_f64x4, centroid_b_x_f64x4);
+        __m256d centered_b_x_high_f64x4 = _mm256_sub_pd(b_x_high_f64x4, centroid_b_x_f64x4);
+        __m256d centered_b_y_low_f64x4 = _mm256_sub_pd(b_y_low_f64x4, centroid_b_y_f64x4);
+        __m256d centered_b_y_high_f64x4 = _mm256_sub_pd(b_y_high_f64x4, centroid_b_y_f64x4);
+        __m256d centered_b_z_low_f64x4 = _mm256_sub_pd(b_z_low_f64x4, centroid_b_z_f64x4);
+        __m256d centered_b_z_high_f64x4 = _mm256_sub_pd(b_z_high_f64x4, centroid_b_z_f64x4);
+
+        __m256d rotated_a_x_low_f64x4 = _mm256_fmadd_pd(
+            scaled_rotation_x_z_f64x4, centered_a_z_low_f64x4,
+            _mm256_fmadd_pd(scaled_rotation_x_y_f64x4, centered_a_y_low_f64x4,
+                            _mm256_mul_pd(scaled_rotation_x_x_f64x4, centered_a_x_low_f64x4)));
+        __m256d rotated_a_x_high_f64x4 = _mm256_fmadd_pd(
+            scaled_rotation_x_z_f64x4, centered_a_z_high_f64x4,
+            _mm256_fmadd_pd(scaled_rotation_x_y_f64x4, centered_a_y_high_f64x4,
+                            _mm256_mul_pd(scaled_rotation_x_x_f64x4, centered_a_x_high_f64x4)));
+        __m256d rotated_a_y_low_f64x4 = _mm256_fmadd_pd(
+            scaled_rotation_y_z_f64x4, centered_a_z_low_f64x4,
+            _mm256_fmadd_pd(scaled_rotation_y_y_f64x4, centered_a_y_low_f64x4,
+                            _mm256_mul_pd(scaled_rotation_y_x_f64x4, centered_a_x_low_f64x4)));
+        __m256d rotated_a_y_high_f64x4 = _mm256_fmadd_pd(
+            scaled_rotation_y_z_f64x4, centered_a_z_high_f64x4,
+            _mm256_fmadd_pd(scaled_rotation_y_y_f64x4, centered_a_y_high_f64x4,
+                            _mm256_mul_pd(scaled_rotation_y_x_f64x4, centered_a_x_high_f64x4)));
+        __m256d rotated_a_z_low_f64x4 = _mm256_fmadd_pd(
+            scaled_rotation_z_z_f64x4, centered_a_z_low_f64x4,
+            _mm256_fmadd_pd(scaled_rotation_z_y_f64x4, centered_a_y_low_f64x4,
+                            _mm256_mul_pd(scaled_rotation_z_x_f64x4, centered_a_x_low_f64x4)));
+        __m256d rotated_a_z_high_f64x4 = _mm256_fmadd_pd(
+            scaled_rotation_z_z_f64x4, centered_a_z_high_f64x4,
+            _mm256_fmadd_pd(scaled_rotation_z_y_f64x4, centered_a_y_high_f64x4,
+                            _mm256_mul_pd(scaled_rotation_z_x_f64x4, centered_a_x_high_f64x4)));
+
+        __m256d delta_x_low_f64x4 = _mm256_sub_pd(rotated_a_x_low_f64x4, centered_b_x_low_f64x4);
+        __m256d delta_x_high_f64x4 = _mm256_sub_pd(rotated_a_x_high_f64x4, centered_b_x_high_f64x4);
+        __m256d delta_y_low_f64x4 = _mm256_sub_pd(rotated_a_y_low_f64x4, centered_b_y_low_f64x4);
+        __m256d delta_y_high_f64x4 = _mm256_sub_pd(rotated_a_y_high_f64x4, centered_b_y_high_f64x4);
+        __m256d delta_z_low_f64x4 = _mm256_sub_pd(rotated_a_z_low_f64x4, centered_b_z_low_f64x4);
+        __m256d delta_z_high_f64x4 = _mm256_sub_pd(rotated_a_z_high_f64x4, centered_b_z_high_f64x4);
+
+        __m256d batch_sum_squared_f64x4 = _mm256_add_pd(_mm256_mul_pd(delta_x_low_f64x4, delta_x_low_f64x4),
+                                                        _mm256_mul_pd(delta_x_high_f64x4, delta_x_high_f64x4));
+        batch_sum_squared_f64x4 = _mm256_fmadd_pd(delta_y_low_f64x4, delta_y_low_f64x4, batch_sum_squared_f64x4);
+        batch_sum_squared_f64x4 = _mm256_fmadd_pd(delta_y_high_f64x4, delta_y_high_f64x4, batch_sum_squared_f64x4);
+        batch_sum_squared_f64x4 = _mm256_fmadd_pd(delta_z_low_f64x4, delta_z_low_f64x4, batch_sum_squared_f64x4);
+        batch_sum_squared_f64x4 = _mm256_fmadd_pd(delta_z_high_f64x4, delta_z_high_f64x4, batch_sum_squared_f64x4);
+        sum_squared_f64x4 = _mm256_add_pd(sum_squared_f64x4, batch_sum_squared_f64x4);
+    }
+
+    nk_f64_t sum_squared = nk_reduce_add_f64x4_haswell_(sum_squared_f64x4);
+    for (; index < n; ++index) {
+        nk_f64_t centered_a_x = (nk_f64_t)a[index * 3 + 0] - centroid_a_x,
+                 centered_a_y = (nk_f64_t)a[index * 3 + 1] - centroid_a_y,
+                 centered_a_z = (nk_f64_t)a[index * 3 + 2] - centroid_a_z;
+        nk_f64_t centered_b_x = (nk_f64_t)b[index * 3 + 0] - centroid_b_x,
+                 centered_b_y = (nk_f64_t)b[index * 3 + 1] - centroid_b_y,
+                 centered_b_z = (nk_f64_t)b[index * 3 + 2] - centroid_b_z;
+        nk_f64_t rotated_a_x = scale * (r[0] * centered_a_x + r[1] * centered_a_y + r[2] * centered_a_z),
+                 rotated_a_y = scale * (r[3] * centered_a_x + r[4] * centered_a_y + r[5] * centered_a_z),
+                 rotated_a_z = scale * (r[6] * centered_a_x + r[7] * centered_a_y + r[8] * centered_a_z);
+        nk_f64_t delta_x = rotated_a_x - centered_b_x, delta_y = rotated_a_y - centered_b_y,
+                 delta_z = rotated_a_z - centered_b_z;
+        sum_squared += delta_x * delta_x + delta_y * delta_y + delta_z * delta_z;
+    }
+
+    return sum_squared;
+}
+
+/*  Compute sum of squared distances for f64 after applying rotation (and optional scale).
+ *  Rotation matrix, scale and data are all f64 for full precision.
+ */
+NK_INTERNAL nk_f64_t nk_transformed_ssd_f64_haswell_(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n,
+                                                     nk_f64_t const *r, nk_f64_t scale, nk_f64_t centroid_a_x,
+                                                     nk_f64_t centroid_a_y, nk_f64_t centroid_a_z,
+                                                     nk_f64_t centroid_b_x, nk_f64_t centroid_b_y,
+                                                     nk_f64_t centroid_b_z) {
+    // Broadcast scaled rotation matrix elements
+    __m256d scaled_rotation_x_x_f64x4 = _mm256_set1_pd(scale * r[0]);
+    __m256d scaled_rotation_x_y_f64x4 = _mm256_set1_pd(scale * r[1]);
+    __m256d scaled_rotation_x_z_f64x4 = _mm256_set1_pd(scale * r[2]);
+    __m256d scaled_rotation_y_x_f64x4 = _mm256_set1_pd(scale * r[3]);
+    __m256d scaled_rotation_y_y_f64x4 = _mm256_set1_pd(scale * r[4]);
+    __m256d scaled_rotation_y_z_f64x4 = _mm256_set1_pd(scale * r[5]);
+    __m256d scaled_rotation_z_x_f64x4 = _mm256_set1_pd(scale * r[6]);
+    __m256d scaled_rotation_z_y_f64x4 = _mm256_set1_pd(scale * r[7]);
+    __m256d scaled_rotation_z_z_f64x4 = _mm256_set1_pd(scale * r[8]);
+
+    // Broadcast centroids
+    __m256d centroid_a_x_f64x4 = _mm256_set1_pd(centroid_a_x);
+    __m256d centroid_a_y_f64x4 = _mm256_set1_pd(centroid_a_y);
+    __m256d centroid_a_z_f64x4 = _mm256_set1_pd(centroid_a_z);
+    __m256d centroid_b_x_f64x4 = _mm256_set1_pd(centroid_b_x);
+    __m256d centroid_b_y_f64x4 = _mm256_set1_pd(centroid_b_y);
+    __m256d centroid_b_z_f64x4 = _mm256_set1_pd(centroid_b_z);
+
+    __m256d sum_squared_f64x4 = _mm256_setzero_pd();
+    __m256d sum_squared_compensation_f64x4 = _mm256_setzero_pd();
+    __m256d a_x_f64x4, a_y_f64x4, a_z_f64x4, b_x_f64x4, b_y_f64x4, b_z_f64x4;
+    nk_size_t j = 0;
+
+    for (; j + 4 <= n; j += 4) {
+        nk_deinterleave_f64x4_haswell_(a + j * 3, &a_x_f64x4, &a_y_f64x4, &a_z_f64x4);
+        nk_deinterleave_f64x4_haswell_(b + j * 3, &b_x_f64x4, &b_y_f64x4, &b_z_f64x4);
+
+        // Center points
+        __m256d pa_x_f64x4 = _mm256_sub_pd(a_x_f64x4, centroid_a_x_f64x4);
+        __m256d pa_y_f64x4 = _mm256_sub_pd(a_y_f64x4, centroid_a_y_f64x4);
+        __m256d pa_z_f64x4 = _mm256_sub_pd(a_z_f64x4, centroid_a_z_f64x4);
+        __m256d pb_x_f64x4 = _mm256_sub_pd(b_x_f64x4, centroid_b_x_f64x4);
+        __m256d pb_y_f64x4 = _mm256_sub_pd(b_y_f64x4, centroid_b_y_f64x4);
+        __m256d pb_z_f64x4 = _mm256_sub_pd(b_z_f64x4, centroid_b_z_f64x4);
+
+        // Rotate and scale: ra = scale * R * pa
+        __m256d ra_x_f64x4 = _mm256_fmadd_pd(scaled_rotation_x_z_f64x4, pa_z_f64x4,
+                                             _mm256_fmadd_pd(scaled_rotation_x_y_f64x4, pa_y_f64x4,
+                                                             _mm256_mul_pd(scaled_rotation_x_x_f64x4, pa_x_f64x4)));
+        __m256d ra_y_f64x4 = _mm256_fmadd_pd(scaled_rotation_y_z_f64x4, pa_z_f64x4,
+                                             _mm256_fmadd_pd(scaled_rotation_y_y_f64x4, pa_y_f64x4,
+                                                             _mm256_mul_pd(scaled_rotation_y_x_f64x4, pa_x_f64x4)));
+        __m256d ra_z_f64x4 = _mm256_fmadd_pd(scaled_rotation_z_z_f64x4, pa_z_f64x4,
+                                             _mm256_fmadd_pd(scaled_rotation_z_y_f64x4, pa_y_f64x4,
+                                                             _mm256_mul_pd(scaled_rotation_z_x_f64x4, pa_x_f64x4)));
+
+        // Delta and accumulate
+        __m256d delta_x_f64x4 = _mm256_sub_pd(ra_x_f64x4, pb_x_f64x4);
+        __m256d delta_y_f64x4 = _mm256_sub_pd(ra_y_f64x4, pb_y_f64x4);
+        __m256d delta_z_f64x4 = _mm256_sub_pd(ra_z_f64x4, pb_z_f64x4);
+
+        nk_accumulate_square_f64x4_haswell_(&sum_squared_f64x4, &sum_squared_compensation_f64x4, delta_x_f64x4);
+        nk_accumulate_square_f64x4_haswell_(&sum_squared_f64x4, &sum_squared_compensation_f64x4, delta_y_f64x4);
+        nk_accumulate_square_f64x4_haswell_(&sum_squared_f64x4, &sum_squared_compensation_f64x4, delta_z_f64x4);
+    }
+
+    nk_f64_t sum_squared = nk_dot_stable_sum_f64x4_haswell_(sum_squared_f64x4, sum_squared_compensation_f64x4);
+    nk_f64_t sum_squared_compensation = 0.0;
+
+    // Scalar tail
+    for (; j < n; ++j) {
+        nk_f64_t pa_x = a[j * 3 + 0] - centroid_a_x, pa_y = a[j * 3 + 1] - centroid_a_y,
+                 pa_z = a[j * 3 + 2] - centroid_a_z;
+        nk_f64_t pb_x = b[j * 3 + 0] - centroid_b_x, pb_y = b[j * 3 + 1] - centroid_b_y,
+                 pb_z = b[j * 3 + 2] - centroid_b_z;
+        nk_f64_t ra_x = scale * (r[0] * pa_x + r[1] * pa_y + r[2] * pa_z),
+                 ra_y = scale * (r[3] * pa_x + r[4] * pa_y + r[5] * pa_z),
+                 ra_z = scale * (r[6] * pa_x + r[7] * pa_y + r[8] * pa_z);
+
+        nk_f64_t delta_x = ra_x - pb_x, delta_y = ra_y - pb_y, delta_z = ra_z - pb_z;
+        nk_accumulate_square_f64_(&sum_squared, &sum_squared_compensation, delta_x);
+        nk_accumulate_square_f64_(&sum_squared, &sum_squared_compensation, delta_y);
+        nk_accumulate_square_f64_(&sum_squared, &sum_squared_compensation, delta_z);
+    }
+
+    return sum_squared + sum_squared_compensation;
+}
+
+NK_PUBLIC void nk_rmsd_f32_haswell(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+                                   nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f64_t *result) {
+    if (a_centroid) a_centroid[0] = 0, a_centroid[1] = 0, a_centroid[2] = 0;
+    if (b_centroid) b_centroid[0] = 0, b_centroid[1] = 0, b_centroid[2] = 0;
+    if (rotation)
+        rotation[0] = 1, rotation[1] = 0, rotation[2] = 0, rotation[3] = 0, rotation[4] = 1, rotation[5] = 0,
+        rotation[6] = 0, rotation[7] = 0, rotation[8] = 1;
+    if (scale) *scale = 1.0f;
+
+    __m256d sum_squared_f64x4 = _mm256_setzero_pd();
+    __m256 a_x_f32x8, a_y_f32x8, a_z_f32x8, b_x_f32x8, b_y_f32x8, b_z_f32x8;
+    nk_size_t index = 0;
+
+    for (; index + 8 <= n; index += 8) {
+        nk_deinterleave_f32x8_haswell_(a + index * 3, &a_x_f32x8, &a_y_f32x8, &a_z_f32x8),
+            nk_deinterleave_f32x8_haswell_(b + index * 3, &b_x_f32x8, &b_y_f32x8, &b_z_f32x8);
+
+        __m256d delta_x_low_f64x4 = _mm256_sub_pd(_mm256_cvtps_pd(_mm256_castps256_ps128(a_x_f32x8)),
+                                                  _mm256_cvtps_pd(_mm256_castps256_ps128(b_x_f32x8)));
+        __m256d delta_x_high_f64x4 = _mm256_sub_pd(_mm256_cvtps_pd(_mm256_extractf128_ps(a_x_f32x8, 1)),
+                                                   _mm256_cvtps_pd(_mm256_extractf128_ps(b_x_f32x8, 1)));
+        __m256d delta_y_low_f64x4 = _mm256_sub_pd(_mm256_cvtps_pd(_mm256_castps256_ps128(a_y_f32x8)),
+                                                  _mm256_cvtps_pd(_mm256_castps256_ps128(b_y_f32x8)));
+        __m256d delta_y_high_f64x4 = _mm256_sub_pd(_mm256_cvtps_pd(_mm256_extractf128_ps(a_y_f32x8, 1)),
+                                                   _mm256_cvtps_pd(_mm256_extractf128_ps(b_y_f32x8, 1)));
+        __m256d delta_z_low_f64x4 = _mm256_sub_pd(_mm256_cvtps_pd(_mm256_castps256_ps128(a_z_f32x8)),
+                                                  _mm256_cvtps_pd(_mm256_castps256_ps128(b_z_f32x8)));
+        __m256d delta_z_high_f64x4 = _mm256_sub_pd(_mm256_cvtps_pd(_mm256_extractf128_ps(a_z_f32x8, 1)),
+                                                   _mm256_cvtps_pd(_mm256_extractf128_ps(b_z_f32x8, 1)));
+
+        __m256d batch_sum_squared_f64x4 = _mm256_add_pd(_mm256_mul_pd(delta_x_low_f64x4, delta_x_low_f64x4),
+                                                        _mm256_mul_pd(delta_x_high_f64x4, delta_x_high_f64x4));
+        batch_sum_squared_f64x4 = _mm256_fmadd_pd(delta_y_low_f64x4, delta_y_low_f64x4, batch_sum_squared_f64x4);
+        batch_sum_squared_f64x4 = _mm256_fmadd_pd(delta_y_high_f64x4, delta_y_high_f64x4, batch_sum_squared_f64x4);
+        batch_sum_squared_f64x4 = _mm256_fmadd_pd(delta_z_low_f64x4, delta_z_low_f64x4, batch_sum_squared_f64x4);
+        batch_sum_squared_f64x4 = _mm256_fmadd_pd(delta_z_high_f64x4, delta_z_high_f64x4, batch_sum_squared_f64x4);
+        sum_squared_f64x4 = _mm256_add_pd(sum_squared_f64x4, batch_sum_squared_f64x4);
+    }
+
+    nk_f64_t sum_squared = nk_reduce_add_f64x4_haswell_(sum_squared_f64x4);
+
+    for (; index < n; ++index) {
+        nk_f64_t delta_x = (nk_f64_t)a[index * 3 + 0] - (nk_f64_t)b[index * 3 + 0];
+        nk_f64_t delta_y = (nk_f64_t)a[index * 3 + 1] - (nk_f64_t)b[index * 3 + 1];
+        nk_f64_t delta_z = (nk_f64_t)a[index * 3 + 2] - (nk_f64_t)b[index * 3 + 2];
+        sum_squared += delta_x * delta_x + delta_y * delta_y + delta_z * delta_z;
+    }
+
+    *result = nk_f64_sqrt_haswell(sum_squared / (nk_f64_t)n);
+}
+
+NK_PUBLIC void nk_rmsd_f64_haswell(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *a_centroid,
+                                   nk_f64_t *b_centroid, nk_f64_t *rotation, nk_f64_t *scale, nk_f64_t *result) {
+    if (a_centroid) a_centroid[0] = 0, a_centroid[1] = 0, a_centroid[2] = 0;
+    if (b_centroid) b_centroid[0] = 0, b_centroid[1] = 0, b_centroid[2] = 0;
+    if (rotation)
+        rotation[0] = 1, rotation[1] = 0, rotation[2] = 0, rotation[3] = 0, rotation[4] = 1, rotation[5] = 0,
+        rotation[6] = 0, rotation[7] = 0, rotation[8] = 1;
+    if (scale) *scale = 1.0;
+    __m256d const zeros_f64x4 = _mm256_setzero_pd();
+
+    __m256d sum_squared_x_f64x4 = zeros_f64x4, sum_squared_y_f64x4 = zeros_f64x4, sum_squared_z_f64x4 = zeros_f64x4;
+
+    __m256d a_x_f64x4, a_y_f64x4, a_z_f64x4, b_x_f64x4, b_y_f64x4, b_z_f64x4;
+    nk_size_t i = 0;
+
+    // Main loop with 2× unrolling
+    for (; i + 8 <= n; i += 8) {
+        nk_deinterleave_f64x4_haswell_(a + i * 3, &a_x_f64x4, &a_y_f64x4, &a_z_f64x4);
+        nk_deinterleave_f64x4_haswell_(b + i * 3, &b_x_f64x4, &b_y_f64x4, &b_z_f64x4);
+
+        __m256d delta_x_f64x4 = _mm256_sub_pd(a_x_f64x4, b_x_f64x4);
+        __m256d delta_y_f64x4 = _mm256_sub_pd(a_y_f64x4, b_y_f64x4);
+        __m256d delta_z_f64x4 = _mm256_sub_pd(a_z_f64x4, b_z_f64x4);
+
+        sum_squared_x_f64x4 = _mm256_fmadd_pd(delta_x_f64x4, delta_x_f64x4, sum_squared_x_f64x4);
+        sum_squared_y_f64x4 = _mm256_fmadd_pd(delta_y_f64x4, delta_y_f64x4, sum_squared_y_f64x4);
+        sum_squared_z_f64x4 = _mm256_fmadd_pd(delta_z_f64x4, delta_z_f64x4, sum_squared_z_f64x4);
+
+        __m256d a_x1_f64x4, a_y1_f64x4, a_z1_f64x4, b_x1_f64x4, b_y1_f64x4, b_z1_f64x4;
+        nk_deinterleave_f64x4_haswell_(a + (i + 4) * 3, &a_x1_f64x4, &a_y1_f64x4, &a_z1_f64x4);
+        nk_deinterleave_f64x4_haswell_(b + (i + 4) * 3, &b_x1_f64x4, &b_y1_f64x4, &b_z1_f64x4);
+
+        __m256d delta_x1_f64x4 = _mm256_sub_pd(a_x1_f64x4, b_x1_f64x4);
+        __m256d delta_y1_f64x4 = _mm256_sub_pd(a_y1_f64x4, b_y1_f64x4);
+        __m256d delta_z1_f64x4 = _mm256_sub_pd(a_z1_f64x4, b_z1_f64x4);
+
+        sum_squared_x_f64x4 = _mm256_fmadd_pd(delta_x1_f64x4, delta_x1_f64x4, sum_squared_x_f64x4);
+        sum_squared_y_f64x4 = _mm256_fmadd_pd(delta_y1_f64x4, delta_y1_f64x4, sum_squared_y_f64x4);
+        sum_squared_z_f64x4 = _mm256_fmadd_pd(delta_z1_f64x4, delta_z1_f64x4, sum_squared_z_f64x4);
+    }
+
+    for (; i + 4 <= n; i += 4) {
+        nk_deinterleave_f64x4_haswell_(a + i * 3, &a_x_f64x4, &a_y_f64x4, &a_z_f64x4);
+        nk_deinterleave_f64x4_haswell_(b + i * 3, &b_x_f64x4, &b_y_f64x4, &b_z_f64x4);
+
+        __m256d delta_x_f64x4 = _mm256_sub_pd(a_x_f64x4, b_x_f64x4);
+        __m256d delta_y_f64x4 = _mm256_sub_pd(a_y_f64x4, b_y_f64x4);
+        __m256d delta_z_f64x4 = _mm256_sub_pd(a_z_f64x4, b_z_f64x4);
+
+        sum_squared_x_f64x4 = _mm256_fmadd_pd(delta_x_f64x4, delta_x_f64x4, sum_squared_x_f64x4);
+        sum_squared_y_f64x4 = _mm256_fmadd_pd(delta_y_f64x4, delta_y_f64x4, sum_squared_y_f64x4);
+        sum_squared_z_f64x4 = _mm256_fmadd_pd(delta_z_f64x4, delta_z_f64x4, sum_squared_z_f64x4);
+    }
+
+    nk_f64_t total_sq_x = nk_reduce_stable_f64x4_haswell_(sum_squared_x_f64x4), total_sq_x_compensation = 0.0;
+    nk_f64_t total_sq_y = nk_reduce_stable_f64x4_haswell_(sum_squared_y_f64x4), total_sq_y_compensation = 0.0;
+    nk_f64_t total_sq_z = nk_reduce_stable_f64x4_haswell_(sum_squared_z_f64x4), total_sq_z_compensation = 0.0;
+
+    for (; i < n; ++i) {
+        nk_f64_t delta_x = a[i * 3 + 0] - b[i * 3 + 0];
+        nk_f64_t delta_y = a[i * 3 + 1] - b[i * 3 + 1];
+        nk_f64_t delta_z = a[i * 3 + 2] - b[i * 3 + 2];
+        nk_accumulate_square_f64_(&total_sq_x, &total_sq_x_compensation, delta_x);
+        nk_accumulate_square_f64_(&total_sq_y, &total_sq_y_compensation, delta_y);
+        nk_accumulate_square_f64_(&total_sq_z, &total_sq_z_compensation, delta_z);
+    }
+
+    total_sq_x += total_sq_x_compensation, total_sq_y += total_sq_y_compensation, total_sq_z += total_sq_z_compensation;
+
+    *result = nk_f64_sqrt_haswell((total_sq_x + total_sq_y + total_sq_z) / (nk_f64_t)n);
+}
+
+NK_PUBLIC void nk_kabsch_f32_haswell(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+                                     nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f64_t *result) {
+    if (scale) *scale = 1.0f;
+    __m256d sum_a_x_f64x4 = _mm256_setzero_pd(), sum_a_y_f64x4 = _mm256_setzero_pd();
+    __m256d sum_a_z_f64x4 = _mm256_setzero_pd(), sum_b_x_f64x4 = _mm256_setzero_pd();
+    __m256d sum_b_y_f64x4 = _mm256_setzero_pd(), sum_b_z_f64x4 = _mm256_setzero_pd();
+    __m256d covariance_00_f64x4 = _mm256_setzero_pd(), covariance_01_f64x4 = _mm256_setzero_pd();
+    __m256d covariance_02_f64x4 = _mm256_setzero_pd(), covariance_10_f64x4 = _mm256_setzero_pd();
+    __m256d covariance_11_f64x4 = _mm256_setzero_pd(), covariance_12_f64x4 = _mm256_setzero_pd();
+    __m256d covariance_20_f64x4 = _mm256_setzero_pd(), covariance_21_f64x4 = _mm256_setzero_pd();
+    __m256d covariance_22_f64x4 = _mm256_setzero_pd();
+    __m256 a_x_f32x8, a_y_f32x8, a_z_f32x8, b_x_f32x8, b_y_f32x8, b_z_f32x8;
+    nk_size_t index = 0;
+
+    for (; index + 8 <= n; index += 8) {
+        nk_deinterleave_f32x8_haswell_(a + index * 3, &a_x_f32x8, &a_y_f32x8, &a_z_f32x8),
+            nk_deinterleave_f32x8_haswell_(b + index * 3, &b_x_f32x8, &b_y_f32x8, &b_z_f32x8);
+        __m256d a_x_low_f64x4 = _mm256_cvtps_pd(_mm256_castps256_ps128(a_x_f32x8));
+        __m256d a_x_high_f64x4 = _mm256_cvtps_pd(_mm256_extractf128_ps(a_x_f32x8, 1));
+        __m256d a_y_low_f64x4 = _mm256_cvtps_pd(_mm256_castps256_ps128(a_y_f32x8));
+        __m256d a_y_high_f64x4 = _mm256_cvtps_pd(_mm256_extractf128_ps(a_y_f32x8, 1));
+        __m256d a_z_low_f64x4 = _mm256_cvtps_pd(_mm256_castps256_ps128(a_z_f32x8));
+        __m256d a_z_high_f64x4 = _mm256_cvtps_pd(_mm256_extractf128_ps(a_z_f32x8, 1));
+        __m256d b_x_low_f64x4 = _mm256_cvtps_pd(_mm256_castps256_ps128(b_x_f32x8));
+        __m256d b_x_high_f64x4 = _mm256_cvtps_pd(_mm256_extractf128_ps(b_x_f32x8, 1));
+        __m256d b_y_low_f64x4 = _mm256_cvtps_pd(_mm256_castps256_ps128(b_y_f32x8));
+        __m256d b_y_high_f64x4 = _mm256_cvtps_pd(_mm256_extractf128_ps(b_y_f32x8, 1));
+        __m256d b_z_low_f64x4 = _mm256_cvtps_pd(_mm256_castps256_ps128(b_z_f32x8));
+        __m256d b_z_high_f64x4 = _mm256_cvtps_pd(_mm256_extractf128_ps(b_z_f32x8, 1));
+
+        sum_a_x_f64x4 = _mm256_add_pd(sum_a_x_f64x4, _mm256_add_pd(a_x_low_f64x4, a_x_high_f64x4));
+        sum_a_y_f64x4 = _mm256_add_pd(sum_a_y_f64x4, _mm256_add_pd(a_y_low_f64x4, a_y_high_f64x4));
+        sum_a_z_f64x4 = _mm256_add_pd(sum_a_z_f64x4, _mm256_add_pd(a_z_low_f64x4, a_z_high_f64x4));
+        sum_b_x_f64x4 = _mm256_add_pd(sum_b_x_f64x4, _mm256_add_pd(b_x_low_f64x4, b_x_high_f64x4));
+        sum_b_y_f64x4 = _mm256_add_pd(sum_b_y_f64x4, _mm256_add_pd(b_y_low_f64x4, b_y_high_f64x4));
+        sum_b_z_f64x4 = _mm256_add_pd(sum_b_z_f64x4, _mm256_add_pd(b_z_low_f64x4, b_z_high_f64x4));
+
+        covariance_00_f64x4 = _mm256_add_pd(
+            covariance_00_f64x4,
+            _mm256_add_pd(_mm256_mul_pd(a_x_low_f64x4, b_x_low_f64x4), _mm256_mul_pd(a_x_high_f64x4, b_x_high_f64x4)));
+        covariance_01_f64x4 = _mm256_add_pd(
+            covariance_01_f64x4,
+            _mm256_add_pd(_mm256_mul_pd(a_x_low_f64x4, b_y_low_f64x4), _mm256_mul_pd(a_x_high_f64x4, b_y_high_f64x4)));
+        covariance_02_f64x4 = _mm256_add_pd(
+            covariance_02_f64x4,
+            _mm256_add_pd(_mm256_mul_pd(a_x_low_f64x4, b_z_low_f64x4), _mm256_mul_pd(a_x_high_f64x4, b_z_high_f64x4)));
+        covariance_10_f64x4 = _mm256_add_pd(
+            covariance_10_f64x4,
+            _mm256_add_pd(_mm256_mul_pd(a_y_low_f64x4, b_x_low_f64x4), _mm256_mul_pd(a_y_high_f64x4, b_x_high_f64x4)));
+        covariance_11_f64x4 = _mm256_add_pd(
+            covariance_11_f64x4,
+            _mm256_add_pd(_mm256_mul_pd(a_y_low_f64x4, b_y_low_f64x4), _mm256_mul_pd(a_y_high_f64x4, b_y_high_f64x4)));
+        covariance_12_f64x4 = _mm256_add_pd(
+            covariance_12_f64x4,
+            _mm256_add_pd(_mm256_mul_pd(a_y_low_f64x4, b_z_low_f64x4), _mm256_mul_pd(a_y_high_f64x4, b_z_high_f64x4)));
+        covariance_20_f64x4 = _mm256_add_pd(
+            covariance_20_f64x4,
+            _mm256_add_pd(_mm256_mul_pd(a_z_low_f64x4, b_x_low_f64x4), _mm256_mul_pd(a_z_high_f64x4, b_x_high_f64x4)));
+        covariance_21_f64x4 = _mm256_add_pd(
+            covariance_21_f64x4,
+            _mm256_add_pd(_mm256_mul_pd(a_z_low_f64x4, b_y_low_f64x4), _mm256_mul_pd(a_z_high_f64x4, b_y_high_f64x4)));
+        covariance_22_f64x4 = _mm256_add_pd(
+            covariance_22_f64x4,
+            _mm256_add_pd(_mm256_mul_pd(a_z_low_f64x4, b_z_low_f64x4), _mm256_mul_pd(a_z_high_f64x4, b_z_high_f64x4)));
+    }
+
+    nk_f64_t sum_a_x = nk_reduce_add_f64x4_haswell_(sum_a_x_f64x4);
+    nk_f64_t sum_a_y = nk_reduce_add_f64x4_haswell_(sum_a_y_f64x4);
+    nk_f64_t sum_a_z = nk_reduce_add_f64x4_haswell_(sum_a_z_f64x4);
+    nk_f64_t sum_b_x = nk_reduce_add_f64x4_haswell_(sum_b_x_f64x4);
+    nk_f64_t sum_b_y = nk_reduce_add_f64x4_haswell_(sum_b_y_f64x4);
+    nk_f64_t sum_b_z = nk_reduce_add_f64x4_haswell_(sum_b_z_f64x4);
+    nk_f64_t h[9] = {
+        nk_reduce_add_f64x4_haswell_(covariance_00_f64x4), nk_reduce_add_f64x4_haswell_(covariance_01_f64x4),
+        nk_reduce_add_f64x4_haswell_(covariance_02_f64x4), nk_reduce_add_f64x4_haswell_(covariance_10_f64x4),
+        nk_reduce_add_f64x4_haswell_(covariance_11_f64x4), nk_reduce_add_f64x4_haswell_(covariance_12_f64x4),
+        nk_reduce_add_f64x4_haswell_(covariance_20_f64x4), nk_reduce_add_f64x4_haswell_(covariance_21_f64x4),
+        nk_reduce_add_f64x4_haswell_(covariance_22_f64x4)};
+
+    for (; index < n; ++index) {
+        nk_f64_t a_x = a[index * 3 + 0], a_y = a[index * 3 + 1], a_z = a[index * 3 + 2];
+        nk_f64_t b_x = b[index * 3 + 0], b_y = b[index * 3 + 1], b_z = b[index * 3 + 2];
+        sum_a_x += a_x, sum_a_y += a_y, sum_a_z += a_z;
+        sum_b_x += b_x, sum_b_y += b_y, sum_b_z += b_z;
+        h[0] += a_x * b_x, h[1] += a_x * b_y, h[2] += a_x * b_z;
+        h[3] += a_y * b_x, h[4] += a_y * b_y, h[5] += a_y * b_z;
+        h[6] += a_z * b_x, h[7] += a_z * b_y, h[8] += a_z * b_z;
+    }
+
+    nk_f64_t inv_n = 1.0 / (nk_f64_t)n;
+    nk_f64_t centroid_a_x = sum_a_x * inv_n, centroid_a_y = sum_a_y * inv_n, centroid_a_z = sum_a_z * inv_n;
+    nk_f64_t centroid_b_x = sum_b_x * inv_n, centroid_b_y = sum_b_y * inv_n, centroid_b_z = sum_b_z * inv_n;
+    if (a_centroid)
+        a_centroid[0] = (nk_f32_t)centroid_a_x, a_centroid[1] = (nk_f32_t)centroid_a_y,
+        a_centroid[2] = (nk_f32_t)centroid_a_z;
+    if (b_centroid)
+        b_centroid[0] = (nk_f32_t)centroid_b_x, b_centroid[1] = (nk_f32_t)centroid_b_y,
+        b_centroid[2] = (nk_f32_t)centroid_b_z;
+
+    h[0] -= (nk_f64_t)n * centroid_a_x * centroid_b_x, h[1] -= (nk_f64_t)n * centroid_a_x * centroid_b_y,
+        h[2] -= (nk_f64_t)n * centroid_a_x * centroid_b_z, h[3] -= (nk_f64_t)n * centroid_a_y * centroid_b_x,
+        h[4] -= (nk_f64_t)n * centroid_a_y * centroid_b_y, h[5] -= (nk_f64_t)n * centroid_a_y * centroid_b_z,
+        h[6] -= (nk_f64_t)n * centroid_a_z * centroid_b_x, h[7] -= (nk_f64_t)n * centroid_a_z * centroid_b_y,
+        h[8] -= (nk_f64_t)n * centroid_a_z * centroid_b_z;
+
+    nk_f64_t cross_covariance[9] = {h[0], h[1], h[2], h[3], h[4], h[5], h[6], h[7], h[8]};
+    nk_f64_t svd_u[9], svd_s[9], svd_v[9], r[9];
+    nk_svd3x3_f64_(cross_covariance, svd_u, svd_s, svd_v);
+    r[0] = svd_v[0] * svd_u[0] + svd_v[1] * svd_u[1] + svd_v[2] * svd_u[2];
+    r[1] = svd_v[0] * svd_u[3] + svd_v[1] * svd_u[4] + svd_v[2] * svd_u[5];
+    r[2] = svd_v[0] * svd_u[6] + svd_v[1] * svd_u[7] + svd_v[2] * svd_u[8];
+    r[3] = svd_v[3] * svd_u[0] + svd_v[4] * svd_u[1] + svd_v[5] * svd_u[2];
+    r[4] = svd_v[3] * svd_u[3] + svd_v[4] * svd_u[4] + svd_v[5] * svd_u[5];
+    r[5] = svd_v[3] * svd_u[6] + svd_v[4] * svd_u[7] + svd_v[5] * svd_u[8];
+    r[6] = svd_v[6] * svd_u[0] + svd_v[7] * svd_u[1] + svd_v[8] * svd_u[2];
+    r[7] = svd_v[6] * svd_u[3] + svd_v[7] * svd_u[4] + svd_v[8] * svd_u[5];
+    r[8] = svd_v[6] * svd_u[6] + svd_v[7] * svd_u[7] + svd_v[8] * svd_u[8];
+    if (nk_det3x3_f64_(r) < 0) {
+        svd_v[2] = -svd_v[2], svd_v[5] = -svd_v[5], svd_v[8] = -svd_v[8];
+        r[0] = svd_v[0] * svd_u[0] + svd_v[1] * svd_u[1] + svd_v[2] * svd_u[2];
+        r[1] = svd_v[0] * svd_u[3] + svd_v[1] * svd_u[4] + svd_v[2] * svd_u[5];
+        r[2] = svd_v[0] * svd_u[6] + svd_v[1] * svd_u[7] + svd_v[2] * svd_u[8];
+        r[3] = svd_v[3] * svd_u[0] + svd_v[4] * svd_u[1] + svd_v[5] * svd_u[2];
+        r[4] = svd_v[3] * svd_u[3] + svd_v[4] * svd_u[4] + svd_v[5] * svd_u[5];
+        r[5] = svd_v[3] * svd_u[6] + svd_v[4] * svd_u[7] + svd_v[5] * svd_u[8];
+        r[6] = svd_v[6] * svd_u[0] + svd_v[7] * svd_u[1] + svd_v[8] * svd_u[2];
+        r[7] = svd_v[6] * svd_u[3] + svd_v[7] * svd_u[4] + svd_v[8] * svd_u[5];
+        r[8] = svd_v[6] * svd_u[6] + svd_v[7] * svd_u[7] + svd_v[8] * svd_u[8];
+    }
+
+    if (rotation)
+        for (int j = 0; j != 9; ++j) rotation[j] = (nk_f32_t)r[j];
+    nk_f64_t sum_squared = nk_transformed_ssd_f32_haswell_(a, b, n, r, 1.0, centroid_a_x, centroid_a_y, centroid_a_z,
+                                                           centroid_b_x, centroid_b_y, centroid_b_z);
+    *result = nk_f64_sqrt_haswell(sum_squared / (nk_f64_t)n);
+}
+
+NK_PUBLIC void nk_kabsch_f64_haswell(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *a_centroid,
+                                     nk_f64_t *b_centroid, nk_f64_t *rotation, nk_f64_t *scale, nk_f64_t *result) {
+    __m256d const zeros_f64x4 = _mm256_setzero_pd();
+
+    // Accumulators for centroids
+    __m256d sum_a_x_f64x4 = zeros_f64x4, sum_a_y_f64x4 = zeros_f64x4, sum_a_z_f64x4 = zeros_f64x4;
+    __m256d sum_b_x_f64x4 = zeros_f64x4, sum_b_y_f64x4 = zeros_f64x4, sum_b_z_f64x4 = zeros_f64x4;
+
+    // Accumulators for covariance matrix (sum of outer products)
+    __m256d cov_xx_f64x4 = zeros_f64x4, cov_xy_f64x4 = zeros_f64x4, cov_xz_f64x4 = zeros_f64x4;
+    __m256d cov_yx_f64x4 = zeros_f64x4, cov_yy_f64x4 = zeros_f64x4, cov_yz_f64x4 = zeros_f64x4;
+    __m256d cov_zx_f64x4 = zeros_f64x4, cov_zy_f64x4 = zeros_f64x4, cov_zz_f64x4 = zeros_f64x4;
+
+    nk_size_t i = 0;
+    __m256d a_x_f64x4, a_y_f64x4, a_z_f64x4, b_x_f64x4, b_y_f64x4, b_z_f64x4;
+
+    // Fused single-pass
+    for (; i + 4 <= n; i += 4) {
+        nk_deinterleave_f64x4_haswell_(a + i * 3, &a_x_f64x4, &a_y_f64x4, &a_z_f64x4);
+        nk_deinterleave_f64x4_haswell_(b + i * 3, &b_x_f64x4, &b_y_f64x4, &b_z_f64x4);
+
+        sum_a_x_f64x4 = _mm256_add_pd(sum_a_x_f64x4, a_x_f64x4);
+        sum_a_y_f64x4 = _mm256_add_pd(sum_a_y_f64x4, a_y_f64x4);
+        sum_a_z_f64x4 = _mm256_add_pd(sum_a_z_f64x4, a_z_f64x4);
+        sum_b_x_f64x4 = _mm256_add_pd(sum_b_x_f64x4, b_x_f64x4);
+        sum_b_y_f64x4 = _mm256_add_pd(sum_b_y_f64x4, b_y_f64x4);
+        sum_b_z_f64x4 = _mm256_add_pd(sum_b_z_f64x4, b_z_f64x4);
+
+        cov_xx_f64x4 = _mm256_fmadd_pd(a_x_f64x4, b_x_f64x4, cov_xx_f64x4);
+        cov_xy_f64x4 = _mm256_fmadd_pd(a_x_f64x4, b_y_f64x4, cov_xy_f64x4);
+        cov_xz_f64x4 = _mm256_fmadd_pd(a_x_f64x4, b_z_f64x4, cov_xz_f64x4);
+        cov_yx_f64x4 = _mm256_fmadd_pd(a_y_f64x4, b_x_f64x4, cov_yx_f64x4);
+        cov_yy_f64x4 = _mm256_fmadd_pd(a_y_f64x4, b_y_f64x4, cov_yy_f64x4);
+        cov_yz_f64x4 = _mm256_fmadd_pd(a_y_f64x4, b_z_f64x4, cov_yz_f64x4);
+        cov_zx_f64x4 = _mm256_fmadd_pd(a_z_f64x4, b_x_f64x4, cov_zx_f64x4);
+        cov_zy_f64x4 = _mm256_fmadd_pd(a_z_f64x4, b_y_f64x4, cov_zy_f64x4);
+        cov_zz_f64x4 = _mm256_fmadd_pd(a_z_f64x4, b_z_f64x4, cov_zz_f64x4);
+    }
+
+    // Reduce vector accumulators
+    nk_f64_t sum_a_x = nk_reduce_stable_f64x4_haswell_(sum_a_x_f64x4), sum_a_x_compensation = 0.0;
+    nk_f64_t sum_a_y = nk_reduce_stable_f64x4_haswell_(sum_a_y_f64x4), sum_a_y_compensation = 0.0;
+    nk_f64_t sum_a_z = nk_reduce_stable_f64x4_haswell_(sum_a_z_f64x4), sum_a_z_compensation = 0.0;
+    nk_f64_t sum_b_x = nk_reduce_stable_f64x4_haswell_(sum_b_x_f64x4), sum_b_x_compensation = 0.0;
+    nk_f64_t sum_b_y = nk_reduce_stable_f64x4_haswell_(sum_b_y_f64x4), sum_b_y_compensation = 0.0;
+    nk_f64_t sum_b_z = nk_reduce_stable_f64x4_haswell_(sum_b_z_f64x4), sum_b_z_compensation = 0.0;
+
+    nk_f64_t covariance_x_x = nk_reduce_stable_f64x4_haswell_(cov_xx_f64x4), covariance_x_x_compensation = 0.0;
+    nk_f64_t covariance_x_y = nk_reduce_stable_f64x4_haswell_(cov_xy_f64x4), covariance_x_y_compensation = 0.0;
+    nk_f64_t covariance_x_z = nk_reduce_stable_f64x4_haswell_(cov_xz_f64x4), covariance_x_z_compensation = 0.0;
+    nk_f64_t covariance_y_x = nk_reduce_stable_f64x4_haswell_(cov_yx_f64x4), covariance_y_x_compensation = 0.0;
+    nk_f64_t covariance_y_y = nk_reduce_stable_f64x4_haswell_(cov_yy_f64x4), covariance_y_y_compensation = 0.0;
+    nk_f64_t covariance_y_z = nk_reduce_stable_f64x4_haswell_(cov_yz_f64x4), covariance_y_z_compensation = 0.0;
+    nk_f64_t covariance_z_x = nk_reduce_stable_f64x4_haswell_(cov_zx_f64x4), covariance_z_x_compensation = 0.0;
+    nk_f64_t covariance_z_y = nk_reduce_stable_f64x4_haswell_(cov_zy_f64x4), covariance_z_y_compensation = 0.0;
+    nk_f64_t covariance_z_z = nk_reduce_stable_f64x4_haswell_(cov_zz_f64x4), covariance_z_z_compensation = 0.0;
+
+    // Scalar tail
+    for (; i < n; ++i) {
+        nk_f64_t ax = a[i * 3 + 0], ay = a[i * 3 + 1], az = a[i * 3 + 2];
+        nk_f64_t bx = b[i * 3 + 0], by = b[i * 3 + 1], bz = b[i * 3 + 2];
+        nk_accumulate_sum_f64_(&sum_a_x, &sum_a_x_compensation, ax);
+        nk_accumulate_sum_f64_(&sum_a_y, &sum_a_y_compensation, ay);
+        nk_accumulate_sum_f64_(&sum_a_z, &sum_a_z_compensation, az);
+        nk_accumulate_sum_f64_(&sum_b_x, &sum_b_x_compensation, bx);
+        nk_accumulate_sum_f64_(&sum_b_y, &sum_b_y_compensation, by);
+        nk_accumulate_sum_f64_(&sum_b_z, &sum_b_z_compensation, bz);
+        nk_accumulate_product_f64_(&covariance_x_x, &covariance_x_x_compensation, ax, bx);
+        nk_accumulate_product_f64_(&covariance_x_y, &covariance_x_y_compensation, ax, by);
+        nk_accumulate_product_f64_(&covariance_x_z, &covariance_x_z_compensation, ax, bz);
+        nk_accumulate_product_f64_(&covariance_y_x, &covariance_y_x_compensation, ay, bx);
+        nk_accumulate_product_f64_(&covariance_y_y, &covariance_y_y_compensation, ay, by);
+        nk_accumulate_product_f64_(&covariance_y_z, &covariance_y_z_compensation, ay, bz);
+        nk_accumulate_product_f64_(&covariance_z_x, &covariance_z_x_compensation, az, bx);
+        nk_accumulate_product_f64_(&covariance_z_y, &covariance_z_y_compensation, az, by);
+        nk_accumulate_product_f64_(&covariance_z_z, &covariance_z_z_compensation, az, bz);
+    }
+
+    sum_a_x += sum_a_x_compensation, sum_a_y += sum_a_y_compensation, sum_a_z += sum_a_z_compensation;
+    sum_b_x += sum_b_x_compensation, sum_b_y += sum_b_y_compensation, sum_b_z += sum_b_z_compensation;
+    covariance_x_x += covariance_x_x_compensation, covariance_x_y += covariance_x_y_compensation,
+        covariance_x_z += covariance_x_z_compensation;
+    covariance_y_x += covariance_y_x_compensation, covariance_y_y += covariance_y_y_compensation,
+        covariance_y_z += covariance_y_z_compensation;
+    covariance_z_x += covariance_z_x_compensation, covariance_z_y += covariance_z_y_compensation,
+        covariance_z_z += covariance_z_z_compensation;
+
+    // Compute centroids
+    nk_f64_t inv_n = 1.0 / (nk_f64_t)n;
+    nk_f64_t centroid_a_x = sum_a_x * inv_n;
+    nk_f64_t centroid_a_y = sum_a_y * inv_n;
+    nk_f64_t centroid_a_z = sum_a_z * inv_n;
+    nk_f64_t centroid_b_x = sum_b_x * inv_n;
+    nk_f64_t centroid_b_y = sum_b_y * inv_n;
+    nk_f64_t centroid_b_z = sum_b_z * inv_n;
+
+    if (a_centroid) a_centroid[0] = centroid_a_x, a_centroid[1] = centroid_a_y, a_centroid[2] = centroid_a_z;
+    if (b_centroid) b_centroid[0] = centroid_b_x, b_centroid[1] = centroid_b_y, b_centroid[2] = centroid_b_z;
+
+    // Apply centering correction: H_centered = H - n * centroid_a * centroid_bᵀ
+    covariance_x_x -= (nk_f64_t)n * centroid_a_x * centroid_b_x;
+    covariance_x_y -= (nk_f64_t)n * centroid_a_x * centroid_b_y;
+    covariance_x_z -= (nk_f64_t)n * centroid_a_x * centroid_b_z;
+    covariance_y_x -= (nk_f64_t)n * centroid_a_y * centroid_b_x;
+    covariance_y_y -= (nk_f64_t)n * centroid_a_y * centroid_b_y;
+    covariance_y_z -= (nk_f64_t)n * centroid_a_y * centroid_b_z;
+    covariance_z_x -= (nk_f64_t)n * centroid_a_z * centroid_b_x;
+    covariance_z_y -= (nk_f64_t)n * centroid_a_z * centroid_b_y;
+    covariance_z_z -= (nk_f64_t)n * centroid_a_z * centroid_b_z;
+
+    // Compute SVD and optimal rotation using f64 precision (svd_s is 9-element diagonal matrix)
+    nk_f64_t cross_covariance[9] = {covariance_x_x, covariance_x_y, covariance_x_z, covariance_y_x, covariance_y_y,
+                                    covariance_y_z, covariance_z_x, covariance_z_y, covariance_z_z};
+    nk_f64_t svd_u[9], svd_s[9], svd_v[9];
+    nk_svd3x3_f64_(cross_covariance, svd_u, svd_s, svd_v);
+
+    nk_f64_t r[9];
+    nk_rotation_from_svd_f64_haswell_(svd_u, svd_v, r);
+
+    // Handle reflection: if det(R) < 0, negate third column of V and recompute R
+    if (nk_det3x3_f64_(r) < 0) {
+        svd_v[2] = -svd_v[2], svd_v[5] = -svd_v[5], svd_v[8] = -svd_v[8];
+        nk_rotation_from_svd_f64_haswell_(svd_u, svd_v, r);
+    }
+
+    // Output rotation matrix and scale=1.0
+    if (rotation)
+        for (int j = 0; j < 9; ++j) rotation[j] = r[j];
+    if (scale) *scale = 1.0;
+
+    // Compute RMSD after optimal rotation
+    nk_f64_t sum_squared = nk_transformed_ssd_f64_haswell_(a, b, n, r, 1.0, centroid_a_x, centroid_a_y, centroid_a_z,
+                                                           centroid_b_x, centroid_b_y, centroid_b_z);
+    *result = nk_f64_sqrt_haswell(sum_squared * inv_n);
+}
+
+NK_PUBLIC void nk_umeyama_f32_haswell(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+                                      nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f64_t *result) {
+    __m256d sum_a_x_f64x4 = _mm256_setzero_pd(), sum_a_y_f64x4 = _mm256_setzero_pd();
+    __m256d sum_a_z_f64x4 = _mm256_setzero_pd(), sum_b_x_f64x4 = _mm256_setzero_pd();
+    __m256d sum_b_y_f64x4 = _mm256_setzero_pd(), sum_b_z_f64x4 = _mm256_setzero_pd();
+    __m256d covariance_00_f64x4 = _mm256_setzero_pd(), covariance_01_f64x4 = _mm256_setzero_pd();
+    __m256d covariance_02_f64x4 = _mm256_setzero_pd(), covariance_10_f64x4 = _mm256_setzero_pd();
+    __m256d covariance_11_f64x4 = _mm256_setzero_pd(), covariance_12_f64x4 = _mm256_setzero_pd();
+    __m256d covariance_20_f64x4 = _mm256_setzero_pd(), covariance_21_f64x4 = _mm256_setzero_pd();
+    __m256d covariance_22_f64x4 = _mm256_setzero_pd(), variance_a_f64x4 = _mm256_setzero_pd();
+    __m256 a_x_f32x8, a_y_f32x8, a_z_f32x8, b_x_f32x8, b_y_f32x8, b_z_f32x8;
+    nk_size_t index = 0;
+
+    for (; index + 8 <= n; index += 8) {
+        nk_deinterleave_f32x8_haswell_(a + index * 3, &a_x_f32x8, &a_y_f32x8, &a_z_f32x8),
+            nk_deinterleave_f32x8_haswell_(b + index * 3, &b_x_f32x8, &b_y_f32x8, &b_z_f32x8);
+        __m256d a_x_low_f64x4 = _mm256_cvtps_pd(_mm256_castps256_ps128(a_x_f32x8));
+        __m256d a_x_high_f64x4 = _mm256_cvtps_pd(_mm256_extractf128_ps(a_x_f32x8, 1));
+        __m256d a_y_low_f64x4 = _mm256_cvtps_pd(_mm256_castps256_ps128(a_y_f32x8));
+        __m256d a_y_high_f64x4 = _mm256_cvtps_pd(_mm256_extractf128_ps(a_y_f32x8, 1));
+        __m256d a_z_low_f64x4 = _mm256_cvtps_pd(_mm256_castps256_ps128(a_z_f32x8));
+        __m256d a_z_high_f64x4 = _mm256_cvtps_pd(_mm256_extractf128_ps(a_z_f32x8, 1));
+        __m256d b_x_low_f64x4 = _mm256_cvtps_pd(_mm256_castps256_ps128(b_x_f32x8));
+        __m256d b_x_high_f64x4 = _mm256_cvtps_pd(_mm256_extractf128_ps(b_x_f32x8, 1));
+        __m256d b_y_low_f64x4 = _mm256_cvtps_pd(_mm256_castps256_ps128(b_y_f32x8));
+        __m256d b_y_high_f64x4 = _mm256_cvtps_pd(_mm256_extractf128_ps(b_y_f32x8, 1));
+        __m256d b_z_low_f64x4 = _mm256_cvtps_pd(_mm256_castps256_ps128(b_z_f32x8));
+        __m256d b_z_high_f64x4 = _mm256_cvtps_pd(_mm256_extractf128_ps(b_z_f32x8, 1));
+
+        sum_a_x_f64x4 = _mm256_add_pd(sum_a_x_f64x4, _mm256_add_pd(a_x_low_f64x4, a_x_high_f64x4));
+        sum_a_y_f64x4 = _mm256_add_pd(sum_a_y_f64x4, _mm256_add_pd(a_y_low_f64x4, a_y_high_f64x4));
+        sum_a_z_f64x4 = _mm256_add_pd(sum_a_z_f64x4, _mm256_add_pd(a_z_low_f64x4, a_z_high_f64x4));
+        sum_b_x_f64x4 = _mm256_add_pd(sum_b_x_f64x4, _mm256_add_pd(b_x_low_f64x4, b_x_high_f64x4));
+        sum_b_y_f64x4 = _mm256_add_pd(sum_b_y_f64x4, _mm256_add_pd(b_y_low_f64x4, b_y_high_f64x4));
+        sum_b_z_f64x4 = _mm256_add_pd(sum_b_z_f64x4, _mm256_add_pd(b_z_low_f64x4, b_z_high_f64x4));
+        covariance_00_f64x4 = _mm256_add_pd(
+            covariance_00_f64x4,
+            _mm256_add_pd(_mm256_mul_pd(a_x_low_f64x4, b_x_low_f64x4), _mm256_mul_pd(a_x_high_f64x4, b_x_high_f64x4)));
+        covariance_01_f64x4 = _mm256_add_pd(
+            covariance_01_f64x4,
+            _mm256_add_pd(_mm256_mul_pd(a_x_low_f64x4, b_y_low_f64x4), _mm256_mul_pd(a_x_high_f64x4, b_y_high_f64x4)));
+        covariance_02_f64x4 = _mm256_add_pd(
+            covariance_02_f64x4,
+            _mm256_add_pd(_mm256_mul_pd(a_x_low_f64x4, b_z_low_f64x4), _mm256_mul_pd(a_x_high_f64x4, b_z_high_f64x4)));
+        covariance_10_f64x4 = _mm256_add_pd(
+            covariance_10_f64x4,
+            _mm256_add_pd(_mm256_mul_pd(a_y_low_f64x4, b_x_low_f64x4), _mm256_mul_pd(a_y_high_f64x4, b_x_high_f64x4)));
+        covariance_11_f64x4 = _mm256_add_pd(
+            covariance_11_f64x4,
+            _mm256_add_pd(_mm256_mul_pd(a_y_low_f64x4, b_y_low_f64x4), _mm256_mul_pd(a_y_high_f64x4, b_y_high_f64x4)));
+        covariance_12_f64x4 = _mm256_add_pd(
+            covariance_12_f64x4,
+            _mm256_add_pd(_mm256_mul_pd(a_y_low_f64x4, b_z_low_f64x4), _mm256_mul_pd(a_y_high_f64x4, b_z_high_f64x4)));
+        covariance_20_f64x4 = _mm256_add_pd(
+            covariance_20_f64x4,
+            _mm256_add_pd(_mm256_mul_pd(a_z_low_f64x4, b_x_low_f64x4), _mm256_mul_pd(a_z_high_f64x4, b_x_high_f64x4)));
+        covariance_21_f64x4 = _mm256_add_pd(
+            covariance_21_f64x4,
+            _mm256_add_pd(_mm256_mul_pd(a_z_low_f64x4, b_y_low_f64x4), _mm256_mul_pd(a_z_high_f64x4, b_y_high_f64x4)));
+        covariance_22_f64x4 = _mm256_add_pd(
+            covariance_22_f64x4,
+            _mm256_add_pd(_mm256_mul_pd(a_z_low_f64x4, b_z_low_f64x4), _mm256_mul_pd(a_z_high_f64x4, b_z_high_f64x4)));
+        variance_a_f64x4 = _mm256_add_pd(
+            variance_a_f64x4,
+            _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(a_x_low_f64x4, a_x_low_f64x4),
+                                        _mm256_mul_pd(a_x_high_f64x4, a_x_high_f64x4)),
+                          _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(a_y_low_f64x4, a_y_low_f64x4),
+                                                      _mm256_mul_pd(a_y_high_f64x4, a_y_high_f64x4)),
+                                        _mm256_add_pd(_mm256_mul_pd(a_z_low_f64x4, a_z_low_f64x4),
+                                                      _mm256_mul_pd(a_z_high_f64x4, a_z_high_f64x4)))));
+    }
+
+    nk_f64_t sum_a_x = nk_reduce_add_f64x4_haswell_(sum_a_x_f64x4);
+    nk_f64_t sum_a_y = nk_reduce_add_f64x4_haswell_(sum_a_y_f64x4);
+    nk_f64_t sum_a_z = nk_reduce_add_f64x4_haswell_(sum_a_z_f64x4);
+    nk_f64_t sum_b_x = nk_reduce_add_f64x4_haswell_(sum_b_x_f64x4);
+    nk_f64_t sum_b_y = nk_reduce_add_f64x4_haswell_(sum_b_y_f64x4);
+    nk_f64_t sum_b_z = nk_reduce_add_f64x4_haswell_(sum_b_z_f64x4);
+    nk_f64_t h[9] = {
+        nk_reduce_add_f64x4_haswell_(covariance_00_f64x4), nk_reduce_add_f64x4_haswell_(covariance_01_f64x4),
+        nk_reduce_add_f64x4_haswell_(covariance_02_f64x4), nk_reduce_add_f64x4_haswell_(covariance_10_f64x4),
+        nk_reduce_add_f64x4_haswell_(covariance_11_f64x4), nk_reduce_add_f64x4_haswell_(covariance_12_f64x4),
+        nk_reduce_add_f64x4_haswell_(covariance_20_f64x4), nk_reduce_add_f64x4_haswell_(covariance_21_f64x4),
+        nk_reduce_add_f64x4_haswell_(covariance_22_f64x4)};
+    nk_f64_t variance_a = nk_reduce_add_f64x4_haswell_(variance_a_f64x4);
+
+    for (; index < n; ++index) {
+        nk_f64_t a_x = a[index * 3 + 0], a_y = a[index * 3 + 1], a_z = a[index * 3 + 2];
+        nk_f64_t b_x = b[index * 3 + 0], b_y = b[index * 3 + 1], b_z = b[index * 3 + 2];
+        sum_a_x += a_x, sum_a_y += a_y, sum_a_z += a_z;
+        sum_b_x += b_x, sum_b_y += b_y, sum_b_z += b_z;
+        h[0] += a_x * b_x, h[1] += a_x * b_y, h[2] += a_x * b_z;
+        h[3] += a_y * b_x, h[4] += a_y * b_y, h[5] += a_y * b_z;
+        h[6] += a_z * b_x, h[7] += a_z * b_y, h[8] += a_z * b_z;
+        variance_a += a_x * a_x + a_y * a_y + a_z * a_z;
+    }
+
+    nk_f64_t inv_n = 1.0 / (nk_f64_t)n;
+    nk_f64_t centroid_a_x = sum_a_x * inv_n, centroid_a_y = sum_a_y * inv_n, centroid_a_z = sum_a_z * inv_n;
+    nk_f64_t centroid_b_x = sum_b_x * inv_n, centroid_b_y = sum_b_y * inv_n, centroid_b_z = sum_b_z * inv_n;
+    if (a_centroid)
+        a_centroid[0] = (nk_f32_t)centroid_a_x, a_centroid[1] = (nk_f32_t)centroid_a_y,
+        a_centroid[2] = (nk_f32_t)centroid_a_z;
+    if (b_centroid)
+        b_centroid[0] = (nk_f32_t)centroid_b_x, b_centroid[1] = (nk_f32_t)centroid_b_y,
+        b_centroid[2] = (nk_f32_t)centroid_b_z;
+
+    variance_a = variance_a * inv_n -
+                 (centroid_a_x * centroid_a_x + centroid_a_y * centroid_a_y + centroid_a_z * centroid_a_z);
+    h[0] -= (nk_f64_t)n * centroid_a_x * centroid_b_x, h[1] -= (nk_f64_t)n * centroid_a_x * centroid_b_y,
+        h[2] -= (nk_f64_t)n * centroid_a_x * centroid_b_z, h[3] -= (nk_f64_t)n * centroid_a_y * centroid_b_x,
+        h[4] -= (nk_f64_t)n * centroid_a_y * centroid_b_y, h[5] -= (nk_f64_t)n * centroid_a_y * centroid_b_z,
+        h[6] -= (nk_f64_t)n * centroid_a_z * centroid_b_x, h[7] -= (nk_f64_t)n * centroid_a_z * centroid_b_y,
+        h[8] -= (nk_f64_t)n * centroid_a_z * centroid_b_z;
+
+    nk_f64_t cross_covariance[9] = {h[0], h[1], h[2], h[3], h[4], h[5], h[6], h[7], h[8]};
+    nk_f64_t svd_u[9], svd_s[9], svd_v[9], r[9];
+    nk_svd3x3_f64_(cross_covariance, svd_u, svd_s, svd_v);
+    r[0] = svd_v[0] * svd_u[0] + svd_v[1] * svd_u[1] + svd_v[2] * svd_u[2];
+    r[1] = svd_v[0] * svd_u[3] + svd_v[1] * svd_u[4] + svd_v[2] * svd_u[5];
+    r[2] = svd_v[0] * svd_u[6] + svd_v[1] * svd_u[7] + svd_v[2] * svd_u[8];
+    r[3] = svd_v[3] * svd_u[0] + svd_v[4] * svd_u[1] + svd_v[5] * svd_u[2];
+    r[4] = svd_v[3] * svd_u[3] + svd_v[4] * svd_u[4] + svd_v[5] * svd_u[5];
+    r[5] = svd_v[3] * svd_u[6] + svd_v[4] * svd_u[7] + svd_v[5] * svd_u[8];
+    r[6] = svd_v[6] * svd_u[0] + svd_v[7] * svd_u[1] + svd_v[8] * svd_u[2];
+    r[7] = svd_v[6] * svd_u[3] + svd_v[7] * svd_u[4] + svd_v[8] * svd_u[5];
+    r[8] = svd_v[6] * svd_u[6] + svd_v[7] * svd_u[7] + svd_v[8] * svd_u[8];
+
+    nk_f64_t det = nk_det3x3_f64_(r), sign_correction = det < 0 ? -1.0 : 1.0;
+    if (det < 0) {
+        svd_v[2] = -svd_v[2], svd_v[5] = -svd_v[5], svd_v[8] = -svd_v[8];
+        r[0] = svd_v[0] * svd_u[0] + svd_v[1] * svd_u[1] + svd_v[2] * svd_u[2];
+        r[1] = svd_v[0] * svd_u[3] + svd_v[1] * svd_u[4] + svd_v[2] * svd_u[5];
+        r[2] = svd_v[0] * svd_u[6] + svd_v[1] * svd_u[7] + svd_v[2] * svd_u[8];
+        r[3] = svd_v[3] * svd_u[0] + svd_v[4] * svd_u[1] + svd_v[5] * svd_u[2];
+        r[4] = svd_v[3] * svd_u[3] + svd_v[4] * svd_u[4] + svd_v[5] * svd_u[5];
+        r[5] = svd_v[3] * svd_u[6] + svd_v[4] * svd_u[7] + svd_v[5] * svd_u[8];
+        r[6] = svd_v[6] * svd_u[0] + svd_v[7] * svd_u[1] + svd_v[8] * svd_u[2];
+        r[7] = svd_v[6] * svd_u[3] + svd_v[7] * svd_u[4] + svd_v[8] * svd_u[5];
+        r[8] = svd_v[6] * svd_u[6] + svd_v[7] * svd_u[7] + svd_v[8] * svd_u[8];
+    }
+
+    nk_f64_t applied_scale = (svd_s[0] + svd_s[4] + sign_correction * svd_s[8]) / ((nk_f64_t)n * variance_a);
+    if (rotation)
+        for (int j = 0; j != 9; ++j) rotation[j] = (nk_f32_t)r[j];
+    if (scale) *scale = (nk_f32_t)applied_scale;
+    *result = nk_f64_sqrt_haswell(nk_transformed_ssd_f32_haswell_(a, b, n, r, applied_scale, centroid_a_x, centroid_a_y,
+                                                                  centroid_a_z, centroid_b_x, centroid_b_y,
+                                                                  centroid_b_z) /
+                                  (nk_f64_t)n);
+}
+
+NK_PUBLIC void nk_umeyama_f64_haswell(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *a_centroid,
+                                      nk_f64_t *b_centroid, nk_f64_t *rotation, nk_f64_t *scale, nk_f64_t *result) {
+    // Fused single-pass: centroids, covariance, and variance of A
+    __m256d const zeros_f64x4 = _mm256_setzero_pd();
+
+    __m256d sum_a_x_f64x4 = zeros_f64x4, sum_a_y_f64x4 = zeros_f64x4, sum_a_z_f64x4 = zeros_f64x4;
+    __m256d sum_b_x_f64x4 = zeros_f64x4, sum_b_y_f64x4 = zeros_f64x4, sum_b_z_f64x4 = zeros_f64x4;
+    __m256d cov_xx_f64x4 = zeros_f64x4, cov_xy_f64x4 = zeros_f64x4, cov_xz_f64x4 = zeros_f64x4;
+    __m256d cov_yx_f64x4 = zeros_f64x4, cov_yy_f64x4 = zeros_f64x4, cov_yz_f64x4 = zeros_f64x4;
+    __m256d cov_zx_f64x4 = zeros_f64x4, cov_zy_f64x4 = zeros_f64x4, cov_zz_f64x4 = zeros_f64x4;
+    __m256d variance_a_f64x4 = zeros_f64x4;
+
+    nk_size_t i = 0;
+    __m256d a_x_f64x4, a_y_f64x4, a_z_f64x4, b_x_f64x4, b_y_f64x4, b_z_f64x4;
+
+    for (; i + 4 <= n; i += 4) {
+        nk_deinterleave_f64x4_haswell_(a + i * 3, &a_x_f64x4, &a_y_f64x4, &a_z_f64x4);
+        nk_deinterleave_f64x4_haswell_(b + i * 3, &b_x_f64x4, &b_y_f64x4, &b_z_f64x4);
+
+        sum_a_x_f64x4 = _mm256_add_pd(sum_a_x_f64x4, a_x_f64x4),
+        sum_a_y_f64x4 = _mm256_add_pd(sum_a_y_f64x4, a_y_f64x4);
+        sum_a_z_f64x4 = _mm256_add_pd(sum_a_z_f64x4, a_z_f64x4);
+        sum_b_x_f64x4 = _mm256_add_pd(sum_b_x_f64x4, b_x_f64x4),
+        sum_b_y_f64x4 = _mm256_add_pd(sum_b_y_f64x4, b_y_f64x4);
+        sum_b_z_f64x4 = _mm256_add_pd(sum_b_z_f64x4, b_z_f64x4);
+
+        cov_xx_f64x4 = _mm256_fmadd_pd(a_x_f64x4, b_x_f64x4, cov_xx_f64x4),
+        cov_xy_f64x4 = _mm256_fmadd_pd(a_x_f64x4, b_y_f64x4, cov_xy_f64x4);
+        cov_xz_f64x4 = _mm256_fmadd_pd(a_x_f64x4, b_z_f64x4, cov_xz_f64x4);
+        cov_yx_f64x4 = _mm256_fmadd_pd(a_y_f64x4, b_x_f64x4, cov_yx_f64x4),
+        cov_yy_f64x4 = _mm256_fmadd_pd(a_y_f64x4, b_y_f64x4, cov_yy_f64x4);
+        cov_yz_f64x4 = _mm256_fmadd_pd(a_y_f64x4, b_z_f64x4, cov_yz_f64x4);
+        cov_zx_f64x4 = _mm256_fmadd_pd(a_z_f64x4, b_x_f64x4, cov_zx_f64x4),
+        cov_zy_f64x4 = _mm256_fmadd_pd(a_z_f64x4, b_y_f64x4, cov_zy_f64x4);
+        cov_zz_f64x4 = _mm256_fmadd_pd(a_z_f64x4, b_z_f64x4, cov_zz_f64x4);
+        variance_a_f64x4 = _mm256_fmadd_pd(a_x_f64x4, a_x_f64x4, variance_a_f64x4);
+        variance_a_f64x4 = _mm256_fmadd_pd(a_y_f64x4, a_y_f64x4, variance_a_f64x4);
+        variance_a_f64x4 = _mm256_fmadd_pd(a_z_f64x4, a_z_f64x4, variance_a_f64x4);
+    }
+
+    // Reduce vector accumulators
+    nk_f64_t sum_a_x = nk_reduce_stable_f64x4_haswell_(sum_a_x_f64x4), sum_a_x_compensation = 0.0;
+    nk_f64_t sum_a_y = nk_reduce_stable_f64x4_haswell_(sum_a_y_f64x4), sum_a_y_compensation = 0.0;
+    nk_f64_t sum_a_z = nk_reduce_stable_f64x4_haswell_(sum_a_z_f64x4), sum_a_z_compensation = 0.0;
+    nk_f64_t sum_b_x = nk_reduce_stable_f64x4_haswell_(sum_b_x_f64x4), sum_b_x_compensation = 0.0;
+    nk_f64_t sum_b_y = nk_reduce_stable_f64x4_haswell_(sum_b_y_f64x4), sum_b_y_compensation = 0.0;
+    nk_f64_t sum_b_z = nk_reduce_stable_f64x4_haswell_(sum_b_z_f64x4), sum_b_z_compensation = 0.0;
+    nk_f64_t covariance_x_x = nk_reduce_stable_f64x4_haswell_(cov_xx_f64x4), covariance_x_x_compensation = 0.0;
+    nk_f64_t covariance_x_y = nk_reduce_stable_f64x4_haswell_(cov_xy_f64x4), covariance_x_y_compensation = 0.0;
+    nk_f64_t covariance_x_z = nk_reduce_stable_f64x4_haswell_(cov_xz_f64x4), covariance_x_z_compensation = 0.0;
+    nk_f64_t covariance_y_x = nk_reduce_stable_f64x4_haswell_(cov_yx_f64x4), covariance_y_x_compensation = 0.0;
+    nk_f64_t covariance_y_y = nk_reduce_stable_f64x4_haswell_(cov_yy_f64x4), covariance_y_y_compensation = 0.0;
+    nk_f64_t covariance_y_z = nk_reduce_stable_f64x4_haswell_(cov_yz_f64x4), covariance_y_z_compensation = 0.0;
+    nk_f64_t covariance_z_x = nk_reduce_stable_f64x4_haswell_(cov_zx_f64x4), covariance_z_x_compensation = 0.0;
+    nk_f64_t covariance_z_y = nk_reduce_stable_f64x4_haswell_(cov_zy_f64x4), covariance_z_y_compensation = 0.0;
+    nk_f64_t covariance_z_z = nk_reduce_stable_f64x4_haswell_(cov_zz_f64x4), covariance_z_z_compensation = 0.0;
+    nk_f64_t variance_a_sum = nk_reduce_stable_f64x4_haswell_(variance_a_f64x4), variance_a_compensation = 0.0;
+
+    // Scalar tail loop for remaining points
+    for (; i < n; i++) {
+        nk_f64_t ax = a[i * 3 + 0], ay = a[i * 3 + 1], az = a[i * 3 + 2];
+        nk_f64_t bx = b[i * 3 + 0], by = b[i * 3 + 1], bz = b[i * 3 + 2];
+        nk_accumulate_sum_f64_(&sum_a_x, &sum_a_x_compensation, ax);
+        nk_accumulate_sum_f64_(&sum_a_y, &sum_a_y_compensation, ay);
+        nk_accumulate_sum_f64_(&sum_a_z, &sum_a_z_compensation, az);
+        nk_accumulate_sum_f64_(&sum_b_x, &sum_b_x_compensation, bx);
+        nk_accumulate_sum_f64_(&sum_b_y, &sum_b_y_compensation, by);
+        nk_accumulate_sum_f64_(&sum_b_z, &sum_b_z_compensation, bz);
+        nk_accumulate_product_f64_(&covariance_x_x, &covariance_x_x_compensation, ax, bx);
+        nk_accumulate_product_f64_(&covariance_x_y, &covariance_x_y_compensation, ax, by);
+        nk_accumulate_product_f64_(&covariance_x_z, &covariance_x_z_compensation, ax, bz);
+        nk_accumulate_product_f64_(&covariance_y_x, &covariance_y_x_compensation, ay, bx);
+        nk_accumulate_product_f64_(&covariance_y_y, &covariance_y_y_compensation, ay, by);
+        nk_accumulate_product_f64_(&covariance_y_z, &covariance_y_z_compensation, ay, bz);
+        nk_accumulate_product_f64_(&covariance_z_x, &covariance_z_x_compensation, az, bx);
+        nk_accumulate_product_f64_(&covariance_z_y, &covariance_z_y_compensation, az, by);
+        nk_accumulate_product_f64_(&covariance_z_z, &covariance_z_z_compensation, az, bz);
+        nk_accumulate_square_f64_(&variance_a_sum, &variance_a_compensation, ax);
+        nk_accumulate_square_f64_(&variance_a_sum, &variance_a_compensation, ay);
+        nk_accumulate_square_f64_(&variance_a_sum, &variance_a_compensation, az);
+    }
+
+    sum_a_x += sum_a_x_compensation, sum_a_y += sum_a_y_compensation, sum_a_z += sum_a_z_compensation;
+    sum_b_x += sum_b_x_compensation, sum_b_y += sum_b_y_compensation, sum_b_z += sum_b_z_compensation;
+    covariance_x_x += covariance_x_x_compensation, covariance_x_y += covariance_x_y_compensation,
+        covariance_x_z += covariance_x_z_compensation;
+    covariance_y_x += covariance_y_x_compensation, covariance_y_y += covariance_y_y_compensation,
+        covariance_y_z += covariance_y_z_compensation;
+    covariance_z_x += covariance_z_x_compensation, covariance_z_y += covariance_z_y_compensation,
+        covariance_z_z += covariance_z_z_compensation;
+    variance_a_sum += variance_a_compensation;
+
+    // Compute centroids
+    nk_f64_t inv_n = 1.0 / (nk_f64_t)n;
+
+    nk_f64_t centroid_a_x = sum_a_x * inv_n, centroid_a_y = sum_a_y * inv_n, centroid_a_z = sum_a_z * inv_n;
+    nk_f64_t centroid_b_x = sum_b_x * inv_n, centroid_b_y = sum_b_y * inv_n, centroid_b_z = sum_b_z * inv_n;
+
+    if (a_centroid) a_centroid[0] = centroid_a_x, a_centroid[1] = centroid_a_y, a_centroid[2] = centroid_a_z;
+    if (b_centroid) b_centroid[0] = centroid_b_x, b_centroid[1] = centroid_b_y, b_centroid[2] = centroid_b_z;
+
+    // Compute centered covariance and variance
+    nk_f64_t variance_a = variance_a_sum * inv_n -
+                          (centroid_a_x * centroid_a_x + centroid_a_y * centroid_a_y + centroid_a_z * centroid_a_z);
+
+    nk_f64_t cross_covariance[9];
+    cross_covariance[0] = covariance_x_x - sum_a_x * sum_b_x * inv_n;
+    cross_covariance[1] = covariance_x_y - sum_a_x * sum_b_y * inv_n;
+    cross_covariance[2] = covariance_x_z - sum_a_x * sum_b_z * inv_n;
+    cross_covariance[3] = covariance_y_x - sum_a_y * sum_b_x * inv_n;
+    cross_covariance[4] = covariance_y_y - sum_a_y * sum_b_y * inv_n;
+    cross_covariance[5] = covariance_y_z - sum_a_y * sum_b_z * inv_n;
+    cross_covariance[6] = covariance_z_x - sum_a_z * sum_b_x * inv_n;
+    cross_covariance[7] = covariance_z_y - sum_a_z * sum_b_y * inv_n;
+    cross_covariance[8] = covariance_z_z - sum_a_z * sum_b_z * inv_n;
+
+    // SVD using f64 for full precision (svd_s is 9-element diagonal matrix)
+    nk_f64_t svd_u[9], svd_s[9], svd_v[9];
+    nk_svd3x3_f64_(cross_covariance, svd_u, svd_s, svd_v);
+
+    nk_f64_t r[9];
+    nk_rotation_from_svd_f64_haswell_(svd_u, svd_v, r);
+
+    // Scale factor: c = trace(D × S) / (n × variance(a))
+    // svd_s diagonal: [0], [4], [8]
+    nk_f64_t det = nk_det3x3_f64_(r);
+    nk_f64_t d3 = det < 0 ? -1.0 : 1.0;
+    nk_f64_t trace_ds = nk_sum_three_products_f64_(svd_s[0], 1.0, svd_s[4], 1.0, svd_s[8], d3);
+    nk_f64_t c = trace_ds / ((nk_f64_t)n * variance_a);
+    if (scale) *scale = c;
+
+    // Handle reflection
+    if (det < 0) {
+        svd_v[2] = -svd_v[2], svd_v[5] = -svd_v[5], svd_v[8] = -svd_v[8];
+        nk_rotation_from_svd_f64_haswell_(svd_u, svd_v, r);
+    }
+
+    // Output rotation matrix
+    if (rotation)
+        for (int j = 0; j < 9; ++j) rotation[j] = r[j];
+
+    // Compute RMSD with scaling
+    nk_f64_t sum_squared = nk_transformed_ssd_f64_haswell_(a, b, n, r, c, centroid_a_x, centroid_a_y, centroid_a_z,
+                                                           centroid_b_x, centroid_b_y, centroid_b_z);
+    *result = nk_f64_sqrt_haswell(sum_squared * inv_n);
+}
+
+/*  Deinterleave 8 f16 xyz triplets (24 f16 values) and convert to 3 x __m256 f32.
+ *  Uses scalar extraction for clean stride-3 access, then F16C conversion.
+ *
+ *  Input: 24 contiguous f16 [x0,y0,z0, x1,y1,z1, ..., x7,y7,z7]
+ *  Output: x[8], y[8], z[8] vectors in f32
+ */
+NK_INTERNAL void nk_deinterleave_f16x8_to_f32x8_haswell_(nk_f16_t const *ptr, __m256 *x_out, __m256 *y_out,
+                                                         __m256 *z_out) {
+    // Extract x, y, z components with stride-3 access
+    nk_b256_vec_t x_vec, y_vec, z_vec;
+    x_vec.f16s[0] = ptr[0], x_vec.f16s[1] = ptr[3], x_vec.f16s[2] = ptr[6], x_vec.f16s[3] = ptr[9];
+    x_vec.f16s[4] = ptr[12], x_vec.f16s[5] = ptr[15], x_vec.f16s[6] = ptr[18], x_vec.f16s[7] = ptr[21];
+    y_vec.f16s[0] = ptr[1], y_vec.f16s[1] = ptr[4], y_vec.f16s[2] = ptr[7], y_vec.f16s[3] = ptr[10];
+    y_vec.f16s[4] = ptr[13], y_vec.f16s[5] = ptr[16], y_vec.f16s[6] = ptr[19], y_vec.f16s[7] = ptr[22];
+    z_vec.f16s[0] = ptr[2], z_vec.f16s[1] = ptr[5], z_vec.f16s[2] = ptr[8], z_vec.f16s[3] = ptr[11];
+    z_vec.f16s[4] = ptr[14], z_vec.f16s[5] = ptr[17], z_vec.f16s[6] = ptr[20], z_vec.f16s[7] = ptr[23];
+    // Convert f16 to f32 using F16C
+    *x_out = _mm256_cvtph_ps(x_vec.xmms[0]);
+    *y_out = _mm256_cvtph_ps(y_vec.xmms[0]);
+    *z_out = _mm256_cvtph_ps(z_vec.xmms[0]);
+}
+
+/*  Deinterleave 8 bf16 xyz triplets (24 bf16 values) and convert to 3 x __m256 f32.
+ *  Uses scalar extraction for clean stride-3 access, then bit-shift conversion.
+ *
+ *  Input: 24 contiguous bf16 [x0,y0,z0, x1,y1,z1, ..., x7,y7,z7]
+ *  Output: x[8], y[8], z[8] vectors in f32
+ */
+NK_INTERNAL void nk_deinterleave_bf16x8_to_f32x8_haswell_(nk_bf16_t const *ptr, __m256 *x_out, __m256 *y_out,
+                                                          __m256 *z_out) {
+    // Extract x, y, z components with stride-3 access
+    nk_b256_vec_t x_vec, y_vec, z_vec;
+    x_vec.bf16s[0] = ptr[0], x_vec.bf16s[1] = ptr[3], x_vec.bf16s[2] = ptr[6], x_vec.bf16s[3] = ptr[9];
+    x_vec.bf16s[4] = ptr[12], x_vec.bf16s[5] = ptr[15], x_vec.bf16s[6] = ptr[18], x_vec.bf16s[7] = ptr[21];
+    y_vec.bf16s[0] = ptr[1], y_vec.bf16s[1] = ptr[4], y_vec.bf16s[2] = ptr[7], y_vec.bf16s[3] = ptr[10];
+    y_vec.bf16s[4] = ptr[13], y_vec.bf16s[5] = ptr[16], y_vec.bf16s[6] = ptr[19], y_vec.bf16s[7] = ptr[22];
+    z_vec.bf16s[0] = ptr[2], z_vec.bf16s[1] = ptr[5], z_vec.bf16s[2] = ptr[8], z_vec.bf16s[3] = ptr[11];
+    z_vec.bf16s[4] = ptr[14], z_vec.bf16s[5] = ptr[17], z_vec.bf16s[6] = ptr[20], z_vec.bf16s[7] = ptr[23];
+    // Convert bf16 to f32 by left-shifting 16 bits
+    *x_out = nk_bf16x8_to_f32x8_haswell_(x_vec.xmms[0]);
+    *y_out = nk_bf16x8_to_f32x8_haswell_(y_vec.xmms[0]);
+    *z_out = nk_bf16x8_to_f32x8_haswell_(z_vec.xmms[0]);
+}
+
+/*  Compute sum of squared distances for f16 data after applying rotation (and optional scale).
+ *  Loads f16 data, converts to f32 during processing.
+ *  Note: rotation matrix r is f32 (from SVD), scale and computation done in f32.
+ */
+NK_INTERNAL nk_f32_t nk_transformed_ssd_f16_haswell_(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n,
+                                                     nk_f32_t const *r, nk_f32_t scale, nk_f32_t centroid_a_x,
+                                                     nk_f32_t centroid_a_y, nk_f32_t centroid_a_z,
+                                                     nk_f32_t centroid_b_x, nk_f32_t centroid_b_y,
+                                                     nk_f32_t centroid_b_z) {
+    // Broadcast scaled rotation matrix elements
+    __m256 scaled_rotation_x_x_f32x8 = _mm256_set1_ps(scale * r[0]);
+    __m256 scaled_rotation_x_y_f32x8 = _mm256_set1_ps(scale * r[1]);
+    __m256 scaled_rotation_x_z_f32x8 = _mm256_set1_ps(scale * r[2]);
+    __m256 scaled_rotation_y_x_f32x8 = _mm256_set1_ps(scale * r[3]);
+    __m256 scaled_rotation_y_y_f32x8 = _mm256_set1_ps(scale * r[4]);
+    __m256 scaled_rotation_y_z_f32x8 = _mm256_set1_ps(scale * r[5]);
+    __m256 scaled_rotation_z_x_f32x8 = _mm256_set1_ps(scale * r[6]);
+    __m256 scaled_rotation_z_y_f32x8 = _mm256_set1_ps(scale * r[7]);
+    __m256 scaled_rotation_z_z_f32x8 = _mm256_set1_ps(scale * r[8]);
+
+    // Broadcast centroids
+    __m256 centroid_a_x_f32x8 = _mm256_set1_ps(centroid_a_x);
+    __m256 centroid_a_y_f32x8 = _mm256_set1_ps(centroid_a_y);
+    __m256 centroid_a_z_f32x8 = _mm256_set1_ps(centroid_a_z);
+    __m256 centroid_b_x_f32x8 = _mm256_set1_ps(centroid_b_x);
+    __m256 centroid_b_y_f32x8 = _mm256_set1_ps(centroid_b_y);
+    __m256 centroid_b_z_f32x8 = _mm256_set1_ps(centroid_b_z);
+
+    __m256 sum_squared_f32x8 = _mm256_setzero_ps();
+    __m256 a_x_f32x8, a_y_f32x8, a_z_f32x8, b_x_f32x8, b_y_f32x8, b_z_f32x8;
+    nk_size_t j = 0;
+
+    for (; j + 8 <= n; j += 8) {
+        nk_deinterleave_f16x8_to_f32x8_haswell_(a + j * 3, &a_x_f32x8, &a_y_f32x8, &a_z_f32x8);
+        nk_deinterleave_f16x8_to_f32x8_haswell_(b + j * 3, &b_x_f32x8, &b_y_f32x8, &b_z_f32x8);
+
+        // Center points
+        __m256 pa_x_f32x8 = _mm256_sub_ps(a_x_f32x8, centroid_a_x_f32x8);
+        __m256 pa_y_f32x8 = _mm256_sub_ps(a_y_f32x8, centroid_a_y_f32x8);
+        __m256 pa_z_f32x8 = _mm256_sub_ps(a_z_f32x8, centroid_a_z_f32x8);
+        __m256 pb_x_f32x8 = _mm256_sub_ps(b_x_f32x8, centroid_b_x_f32x8);
+        __m256 pb_y_f32x8 = _mm256_sub_ps(b_y_f32x8, centroid_b_y_f32x8);
+        __m256 pb_z_f32x8 = _mm256_sub_ps(b_z_f32x8, centroid_b_z_f32x8);
+
+        // Rotate and scale: ra = scale * R * pa
+        __m256 ra_x_f32x8 = _mm256_fmadd_ps(scaled_rotation_x_z_f32x8, pa_z_f32x8,
+                                            _mm256_fmadd_ps(scaled_rotation_x_y_f32x8, pa_y_f32x8,
+                                                            _mm256_mul_ps(scaled_rotation_x_x_f32x8, pa_x_f32x8)));
+        __m256 ra_y_f32x8 = _mm256_fmadd_ps(scaled_rotation_y_z_f32x8, pa_z_f32x8,
+                                            _mm256_fmadd_ps(scaled_rotation_y_y_f32x8, pa_y_f32x8,
+                                                            _mm256_mul_ps(scaled_rotation_y_x_f32x8, pa_x_f32x8)));
+        __m256 ra_z_f32x8 = _mm256_fmadd_ps(scaled_rotation_z_z_f32x8, pa_z_f32x8,
+                                            _mm256_fmadd_ps(scaled_rotation_z_y_f32x8, pa_y_f32x8,
+                                                            _mm256_mul_ps(scaled_rotation_z_x_f32x8, pa_x_f32x8)));
+
+        // Delta and accumulate
+        __m256 delta_x_f32x8 = _mm256_sub_ps(ra_x_f32x8, pb_x_f32x8);
+        __m256 delta_y_f32x8 = _mm256_sub_ps(ra_y_f32x8, pb_y_f32x8);
+        __m256 delta_z_f32x8 = _mm256_sub_ps(ra_z_f32x8, pb_z_f32x8);
+
+        sum_squared_f32x8 = _mm256_fmadd_ps(delta_x_f32x8, delta_x_f32x8, sum_squared_f32x8);
+        sum_squared_f32x8 = _mm256_fmadd_ps(delta_y_f32x8, delta_y_f32x8, sum_squared_f32x8);
+        sum_squared_f32x8 = _mm256_fmadd_ps(delta_z_f32x8, delta_z_f32x8, sum_squared_f32x8);
+    }
+
+    nk_f32_t sum_squared = nk_reduce_add_f32x8_haswell_(sum_squared_f32x8);
+
+    // Scalar tail
+    for (; j < n; ++j) {
+        nk_f32_t a_x_f32, a_y_f32, a_z_f32, b_x_f32, b_y_f32, b_z_f32;
+        nk_f16_to_f32_haswell(&a[j * 3 + 0], &a_x_f32);
+        nk_f16_to_f32_haswell(&a[j * 3 + 1], &a_y_f32);
+        nk_f16_to_f32_haswell(&a[j * 3 + 2], &a_z_f32);
+        nk_f16_to_f32_haswell(&b[j * 3 + 0], &b_x_f32);
+        nk_f16_to_f32_haswell(&b[j * 3 + 1], &b_y_f32);
+        nk_f16_to_f32_haswell(&b[j * 3 + 2], &b_z_f32);
+
+        nk_f32_t pa_x = a_x_f32 - centroid_a_x, pa_y = a_y_f32 - centroid_a_y, pa_z = a_z_f32 - centroid_a_z;
+        nk_f32_t pb_x = b_x_f32 - centroid_b_x, pb_y = b_y_f32 - centroid_b_y, pb_z = b_z_f32 - centroid_b_z;
+        nk_f32_t ra_x = scale * (r[0] * pa_x + r[1] * pa_y + r[2] * pa_z),
+                 ra_y = scale * (r[3] * pa_x + r[4] * pa_y + r[5] * pa_z),
+                 ra_z = scale * (r[6] * pa_x + r[7] * pa_y + r[8] * pa_z);
+
+        nk_f32_t delta_x = ra_x - pb_x, delta_y = ra_y - pb_y, delta_z = ra_z - pb_z;
+        sum_squared += delta_x * delta_x + delta_y * delta_y + delta_z * delta_z;
+    }
+
+    return sum_squared;
+}
+
+/*  Compute sum of squared distances for bf16 data after applying rotation (and optional scale).
+ *  Loads bf16 data, converts to f32 during processing.
+ *  Note: rotation matrix r is f32 (from SVD), scale and computation done in f32.
+ */
+NK_INTERNAL nk_f32_t nk_transformed_ssd_bf16_haswell_(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n,
+                                                      nk_f32_t const *r, nk_f32_t scale, nk_f32_t centroid_a_x,
+                                                      nk_f32_t centroid_a_y, nk_f32_t centroid_a_z,
+                                                      nk_f32_t centroid_b_x, nk_f32_t centroid_b_y,
+                                                      nk_f32_t centroid_b_z) {
+    // Broadcast scaled rotation matrix elements
+    __m256 scaled_rotation_x_x_f32x8 = _mm256_set1_ps(scale * r[0]);
+    __m256 scaled_rotation_x_y_f32x8 = _mm256_set1_ps(scale * r[1]);
+    __m256 scaled_rotation_x_z_f32x8 = _mm256_set1_ps(scale * r[2]);
+    __m256 scaled_rotation_y_x_f32x8 = _mm256_set1_ps(scale * r[3]);
+    __m256 scaled_rotation_y_y_f32x8 = _mm256_set1_ps(scale * r[4]);
+    __m256 scaled_rotation_y_z_f32x8 = _mm256_set1_ps(scale * r[5]);
+    __m256 scaled_rotation_z_x_f32x8 = _mm256_set1_ps(scale * r[6]);
+    __m256 scaled_rotation_z_y_f32x8 = _mm256_set1_ps(scale * r[7]);
+    __m256 scaled_rotation_z_z_f32x8 = _mm256_set1_ps(scale * r[8]);
+
+    // Broadcast centroids
+    __m256 centroid_a_x_f32x8 = _mm256_set1_ps(centroid_a_x);
+    __m256 centroid_a_y_f32x8 = _mm256_set1_ps(centroid_a_y);
+    __m256 centroid_a_z_f32x8 = _mm256_set1_ps(centroid_a_z);
+    __m256 centroid_b_x_f32x8 = _mm256_set1_ps(centroid_b_x);
+    __m256 centroid_b_y_f32x8 = _mm256_set1_ps(centroid_b_y);
+    __m256 centroid_b_z_f32x8 = _mm256_set1_ps(centroid_b_z);
+
+    __m256 sum_squared_f32x8 = _mm256_setzero_ps();
+    __m256 a_x_f32x8, a_y_f32x8, a_z_f32x8, b_x_f32x8, b_y_f32x8, b_z_f32x8;
+    nk_size_t j = 0;
+
+    for (; j + 8 <= n; j += 8) {
+        nk_deinterleave_bf16x8_to_f32x8_haswell_(a + j * 3, &a_x_f32x8, &a_y_f32x8, &a_z_f32x8);
+        nk_deinterleave_bf16x8_to_f32x8_haswell_(b + j * 3, &b_x_f32x8, &b_y_f32x8, &b_z_f32x8);
+
+        // Center points
+        __m256 pa_x_f32x8 = _mm256_sub_ps(a_x_f32x8, centroid_a_x_f32x8);
+        __m256 pa_y_f32x8 = _mm256_sub_ps(a_y_f32x8, centroid_a_y_f32x8);
+        __m256 pa_z_f32x8 = _mm256_sub_ps(a_z_f32x8, centroid_a_z_f32x8);
+        __m256 pb_x_f32x8 = _mm256_sub_ps(b_x_f32x8, centroid_b_x_f32x8);
+        __m256 pb_y_f32x8 = _mm256_sub_ps(b_y_f32x8, centroid_b_y_f32x8);
+        __m256 pb_z_f32x8 = _mm256_sub_ps(b_z_f32x8, centroid_b_z_f32x8);
+
+        // Rotate and scale: ra = scale * R * pa
+        __m256 ra_x_f32x8 = _mm256_fmadd_ps(scaled_rotation_x_z_f32x8, pa_z_f32x8,
+                                            _mm256_fmadd_ps(scaled_rotation_x_y_f32x8, pa_y_f32x8,
+                                                            _mm256_mul_ps(scaled_rotation_x_x_f32x8, pa_x_f32x8)));
+        __m256 ra_y_f32x8 = _mm256_fmadd_ps(scaled_rotation_y_z_f32x8, pa_z_f32x8,
+                                            _mm256_fmadd_ps(scaled_rotation_y_y_f32x8, pa_y_f32x8,
+                                                            _mm256_mul_ps(scaled_rotation_y_x_f32x8, pa_x_f32x8)));
+        __m256 ra_z_f32x8 = _mm256_fmadd_ps(scaled_rotation_z_z_f32x8, pa_z_f32x8,
+                                            _mm256_fmadd_ps(scaled_rotation_z_y_f32x8, pa_y_f32x8,
+                                                            _mm256_mul_ps(scaled_rotation_z_x_f32x8, pa_x_f32x8)));
+
+        // Delta and accumulate
+        __m256 delta_x_f32x8 = _mm256_sub_ps(ra_x_f32x8, pb_x_f32x8);
+        __m256 delta_y_f32x8 = _mm256_sub_ps(ra_y_f32x8, pb_y_f32x8);
+        __m256 delta_z_f32x8 = _mm256_sub_ps(ra_z_f32x8, pb_z_f32x8);
+
+        sum_squared_f32x8 = _mm256_fmadd_ps(delta_x_f32x8, delta_x_f32x8, sum_squared_f32x8);
+        sum_squared_f32x8 = _mm256_fmadd_ps(delta_y_f32x8, delta_y_f32x8, sum_squared_f32x8);
+        sum_squared_f32x8 = _mm256_fmadd_ps(delta_z_f32x8, delta_z_f32x8, sum_squared_f32x8);
+    }
+
+    nk_f32_t sum_squared = nk_reduce_add_f32x8_haswell_(sum_squared_f32x8);
+
+    // Scalar tail
+    for (; j < n; ++j) {
+        nk_f32_t a_x_f32, a_y_f32, a_z_f32, b_x_f32, b_y_f32, b_z_f32;
+        nk_bf16_to_f32_serial(&a[j * 3 + 0], &a_x_f32);
+        nk_bf16_to_f32_serial(&a[j * 3 + 1], &a_y_f32);
+        nk_bf16_to_f32_serial(&a[j * 3 + 2], &a_z_f32);
+        nk_bf16_to_f32_serial(&b[j * 3 + 0], &b_x_f32);
+        nk_bf16_to_f32_serial(&b[j * 3 + 1], &b_y_f32);
+        nk_bf16_to_f32_serial(&b[j * 3 + 2], &b_z_f32);
+
+        nk_f32_t pa_x = a_x_f32 - centroid_a_x, pa_y = a_y_f32 - centroid_a_y, pa_z = a_z_f32 - centroid_a_z;
+        nk_f32_t pb_x = b_x_f32 - centroid_b_x, pb_y = b_y_f32 - centroid_b_y, pb_z = b_z_f32 - centroid_b_z;
+        nk_f32_t ra_x = scale * (r[0] * pa_x + r[1] * pa_y + r[2] * pa_z),
+                 ra_y = scale * (r[3] * pa_x + r[4] * pa_y + r[5] * pa_z),
+                 ra_z = scale * (r[6] * pa_x + r[7] * pa_y + r[8] * pa_z);
+
+        nk_f32_t delta_x = ra_x - pb_x, delta_y = ra_y - pb_y, delta_z = ra_z - pb_z;
+        sum_squared += delta_x * delta_x + delta_y * delta_y + delta_z * delta_z;
+    }
+
+    return sum_squared;
+}
+
+NK_PUBLIC void nk_rmsd_f16_haswell(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+                                   nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f32_t *result) {
+    // RMSD uses identity rotation and scale=1.0
+    if (rotation)
+        rotation[0] = 1, rotation[1] = 0, rotation[2] = 0, rotation[3] = 0, rotation[4] = 1, rotation[5] = 0,
+        rotation[6] = 0, rotation[7] = 0, rotation[8] = 1;
+    if (scale) *scale = 1.0f;
+
+    __m256 const zeros_f32x8 = _mm256_setzero_ps();
+
+    // Accumulators for centroids and squared differences (all in f32)
+    __m256 sum_a_x_f32x8 = zeros_f32x8, sum_a_y_f32x8 = zeros_f32x8, sum_a_z_f32x8 = zeros_f32x8;
+    __m256 sum_b_x_f32x8 = zeros_f32x8, sum_b_y_f32x8 = zeros_f32x8, sum_b_z_f32x8 = zeros_f32x8;
+    __m256 sum_squared_x_f32x8 = zeros_f32x8, sum_squared_y_f32x8 = zeros_f32x8, sum_squared_z_f32x8 = zeros_f32x8;
+
+    __m256 a_x_f32x8, a_y_f32x8, a_z_f32x8, b_x_f32x8, b_y_f32x8, b_z_f32x8;
+    nk_size_t i = 0;
+
+    // Main loop processing 8 points at a time
+    for (; i + 8 <= n; i += 8) {
+        nk_deinterleave_f16x8_to_f32x8_haswell_(a + i * 3, &a_x_f32x8, &a_y_f32x8, &a_z_f32x8);
+        nk_deinterleave_f16x8_to_f32x8_haswell_(b + i * 3, &b_x_f32x8, &b_y_f32x8, &b_z_f32x8);
+
+        sum_a_x_f32x8 = _mm256_add_ps(sum_a_x_f32x8, a_x_f32x8);
+        sum_a_y_f32x8 = _mm256_add_ps(sum_a_y_f32x8, a_y_f32x8);
+        sum_a_z_f32x8 = _mm256_add_ps(sum_a_z_f32x8, a_z_f32x8);
+        sum_b_x_f32x8 = _mm256_add_ps(sum_b_x_f32x8, b_x_f32x8);
+        sum_b_y_f32x8 = _mm256_add_ps(sum_b_y_f32x8, b_y_f32x8);
+        sum_b_z_f32x8 = _mm256_add_ps(sum_b_z_f32x8, b_z_f32x8);
+
+        __m256 delta_x_f32x8 = _mm256_sub_ps(a_x_f32x8, b_x_f32x8);
+        __m256 delta_y_f32x8 = _mm256_sub_ps(a_y_f32x8, b_y_f32x8);
+        __m256 delta_z_f32x8 = _mm256_sub_ps(a_z_f32x8, b_z_f32x8);
+
+        sum_squared_x_f32x8 = _mm256_fmadd_ps(delta_x_f32x8, delta_x_f32x8, sum_squared_x_f32x8);
+        sum_squared_y_f32x8 = _mm256_fmadd_ps(delta_y_f32x8, delta_y_f32x8, sum_squared_y_f32x8);
+        sum_squared_z_f32x8 = _mm256_fmadd_ps(delta_z_f32x8, delta_z_f32x8, sum_squared_z_f32x8);
+    }
+
+    // Reduce vectors to scalars
+    nk_f32_t total_ax = nk_reduce_add_f32x8_haswell_(sum_a_x_f32x8);
+    nk_f32_t total_ay = nk_reduce_add_f32x8_haswell_(sum_a_y_f32x8);
+    nk_f32_t total_az = nk_reduce_add_f32x8_haswell_(sum_a_z_f32x8);
+    nk_f32_t total_bx = nk_reduce_add_f32x8_haswell_(sum_b_x_f32x8);
+    nk_f32_t total_by = nk_reduce_add_f32x8_haswell_(sum_b_y_f32x8);
+    nk_f32_t total_bz = nk_reduce_add_f32x8_haswell_(sum_b_z_f32x8);
+    nk_f32_t total_sq_x = nk_reduce_add_f32x8_haswell_(sum_squared_x_f32x8);
+    nk_f32_t total_sq_y = nk_reduce_add_f32x8_haswell_(sum_squared_y_f32x8);
+    nk_f32_t total_sq_z = nk_reduce_add_f32x8_haswell_(sum_squared_z_f32x8);
+
+    // Scalar tail
+    for (; i < n; ++i) {
+        nk_f32_t ax, ay, az, bx, by, bz;
+        nk_f16_to_f32_haswell(&a[i * 3 + 0], &ax);
+        nk_f16_to_f32_haswell(&a[i * 3 + 1], &ay);
+        nk_f16_to_f32_haswell(&a[i * 3 + 2], &az);
+        nk_f16_to_f32_haswell(&b[i * 3 + 0], &bx);
+        nk_f16_to_f32_haswell(&b[i * 3 + 1], &by);
+        nk_f16_to_f32_haswell(&b[i * 3 + 2], &bz);
+        total_ax += ax;
+        total_ay += ay;
+        total_az += az;
+        total_bx += bx;
+        total_by += by;
+        total_bz += bz;
+        nk_f32_t delta_x = ax - bx, delta_y = ay - by, delta_z = az - bz;
+        total_sq_x += delta_x * delta_x;
+        total_sq_y += delta_y * delta_y;
+        total_sq_z += delta_z * delta_z;
+    }
+
+    // Compute centroids
+    nk_f32_t inv_n = 1.0f / (nk_f32_t)n;
+    nk_f32_t centroid_a_x = total_ax * inv_n;
+    nk_f32_t centroid_a_y = total_ay * inv_n;
+    nk_f32_t centroid_a_z = total_az * inv_n;
+    nk_f32_t centroid_b_x = total_bx * inv_n;
+    nk_f32_t centroid_b_y = total_by * inv_n;
+    nk_f32_t centroid_b_z = total_bz * inv_n;
+
+    if (a_centroid) a_centroid[0] = centroid_a_x, a_centroid[1] = centroid_a_y, a_centroid[2] = centroid_a_z;
+    if (b_centroid) b_centroid[0] = centroid_b_x, b_centroid[1] = centroid_b_y, b_centroid[2] = centroid_b_z;
+
+    // Compute RMSD
+    nk_f32_t mean_diff_x = centroid_a_x - centroid_b_x;
+    nk_f32_t mean_diff_y = centroid_a_y - centroid_b_y;
+    nk_f32_t mean_diff_z = centroid_a_z - centroid_b_z;
+    nk_f32_t sum_squared = total_sq_x + total_sq_y + total_sq_z;
+    nk_f32_t mean_diff_sq = mean_diff_x * mean_diff_x + mean_diff_y * mean_diff_y + mean_diff_z * mean_diff_z;
+
+    *result = nk_f32_sqrt_haswell(sum_squared * inv_n - mean_diff_sq);
+}
+
+NK_PUBLIC void nk_rmsd_bf16_haswell(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+                                    nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f32_t *result) {
+    // RMSD uses identity rotation and scale=1.0
+    if (rotation)
+        rotation[0] = 1, rotation[1] = 0, rotation[2] = 0, rotation[3] = 0, rotation[4] = 1, rotation[5] = 0,
+        rotation[6] = 0, rotation[7] = 0, rotation[8] = 1;
+    if (scale) *scale = 1.0f;
+
+    __m256 const zeros_f32x8 = _mm256_setzero_ps();
+
+    // Accumulators for centroids and squared differences (all in f32)
+    __m256 sum_a_x_f32x8 = zeros_f32x8, sum_a_y_f32x8 = zeros_f32x8, sum_a_z_f32x8 = zeros_f32x8;
+    __m256 sum_b_x_f32x8 = zeros_f32x8, sum_b_y_f32x8 = zeros_f32x8, sum_b_z_f32x8 = zeros_f32x8;
+    __m256 sum_squared_x_f32x8 = zeros_f32x8, sum_squared_y_f32x8 = zeros_f32x8, sum_squared_z_f32x8 = zeros_f32x8;
+
+    __m256 a_x_f32x8, a_y_f32x8, a_z_f32x8, b_x_f32x8, b_y_f32x8, b_z_f32x8;
+    nk_size_t i = 0;
+
+    // Main loop processing 8 points at a time
+    for (; i + 8 <= n; i += 8) {
+        nk_deinterleave_bf16x8_to_f32x8_haswell_(a + i * 3, &a_x_f32x8, &a_y_f32x8, &a_z_f32x8);
+        nk_deinterleave_bf16x8_to_f32x8_haswell_(b + i * 3, &b_x_f32x8, &b_y_f32x8, &b_z_f32x8);
+
+        sum_a_x_f32x8 = _mm256_add_ps(sum_a_x_f32x8, a_x_f32x8);
+        sum_a_y_f32x8 = _mm256_add_ps(sum_a_y_f32x8, a_y_f32x8);
+        sum_a_z_f32x8 = _mm256_add_ps(sum_a_z_f32x8, a_z_f32x8);
+        sum_b_x_f32x8 = _mm256_add_ps(sum_b_x_f32x8, b_x_f32x8);
+        sum_b_y_f32x8 = _mm256_add_ps(sum_b_y_f32x8, b_y_f32x8);
+        sum_b_z_f32x8 = _mm256_add_ps(sum_b_z_f32x8, b_z_f32x8);
+
+        __m256 delta_x_f32x8 = _mm256_sub_ps(a_x_f32x8, b_x_f32x8);
+        __m256 delta_y_f32x8 = _mm256_sub_ps(a_y_f32x8, b_y_f32x8);
+        __m256 delta_z_f32x8 = _mm256_sub_ps(a_z_f32x8, b_z_f32x8);
+
+        sum_squared_x_f32x8 = _mm256_fmadd_ps(delta_x_f32x8, delta_x_f32x8, sum_squared_x_f32x8);
+        sum_squared_y_f32x8 = _mm256_fmadd_ps(delta_y_f32x8, delta_y_f32x8, sum_squared_y_f32x8);
+        sum_squared_z_f32x8 = _mm256_fmadd_ps(delta_z_f32x8, delta_z_f32x8, sum_squared_z_f32x8);
+    }
+
+    // Reduce vectors to scalars
+    nk_f32_t total_ax = nk_reduce_add_f32x8_haswell_(sum_a_x_f32x8);
+    nk_f32_t total_ay = nk_reduce_add_f32x8_haswell_(sum_a_y_f32x8);
+    nk_f32_t total_az = nk_reduce_add_f32x8_haswell_(sum_a_z_f32x8);
+    nk_f32_t total_bx = nk_reduce_add_f32x8_haswell_(sum_b_x_f32x8);
+    nk_f32_t total_by = nk_reduce_add_f32x8_haswell_(sum_b_y_f32x8);
+    nk_f32_t total_bz = nk_reduce_add_f32x8_haswell_(sum_b_z_f32x8);
+    nk_f32_t total_sq_x = nk_reduce_add_f32x8_haswell_(sum_squared_x_f32x8);
+    nk_f32_t total_sq_y = nk_reduce_add_f32x8_haswell_(sum_squared_y_f32x8);
+    nk_f32_t total_sq_z = nk_reduce_add_f32x8_haswell_(sum_squared_z_f32x8);
+
+    // Scalar tail
+    for (; i < n; ++i) {
+        nk_f32_t ax, ay, az, bx, by, bz;
+        nk_bf16_to_f32_serial(&a[i * 3 + 0], &ax);
+        nk_bf16_to_f32_serial(&a[i * 3 + 1], &ay);
+        nk_bf16_to_f32_serial(&a[i * 3 + 2], &az);
+        nk_bf16_to_f32_serial(&b[i * 3 + 0], &bx);
+        nk_bf16_to_f32_serial(&b[i * 3 + 1], &by);
+        nk_bf16_to_f32_serial(&b[i * 3 + 2], &bz);
+        total_ax += ax;
+        total_ay += ay;
+        total_az += az;
+        total_bx += bx;
+        total_by += by;
+        total_bz += bz;
+        nk_f32_t delta_x = ax - bx, delta_y = ay - by, delta_z = az - bz;
+        total_sq_x += delta_x * delta_x;
+        total_sq_y += delta_y * delta_y;
+        total_sq_z += delta_z * delta_z;
+    }
+
+    // Compute centroids
+    nk_f32_t inv_n = 1.0f / (nk_f32_t)n;
+    nk_f32_t centroid_a_x = total_ax * inv_n;
+    nk_f32_t centroid_a_y = total_ay * inv_n;
+    nk_f32_t centroid_a_z = total_az * inv_n;
+    nk_f32_t centroid_b_x = total_bx * inv_n;
+    nk_f32_t centroid_b_y = total_by * inv_n;
+    nk_f32_t centroid_b_z = total_bz * inv_n;
+
+    if (a_centroid) a_centroid[0] = centroid_a_x, a_centroid[1] = centroid_a_y, a_centroid[2] = centroid_a_z;
+    if (b_centroid) b_centroid[0] = centroid_b_x, b_centroid[1] = centroid_b_y, b_centroid[2] = centroid_b_z;
+
+    // Compute RMSD
+    nk_f32_t mean_diff_x = centroid_a_x - centroid_b_x;
+    nk_f32_t mean_diff_y = centroid_a_y - centroid_b_y;
+    nk_f32_t mean_diff_z = centroid_a_z - centroid_b_z;
+    nk_f32_t sum_squared = total_sq_x + total_sq_y + total_sq_z;
+    nk_f32_t mean_diff_sq = mean_diff_x * mean_diff_x + mean_diff_y * mean_diff_y + mean_diff_z * mean_diff_z;
+
+    *result = nk_f32_sqrt_haswell(sum_squared * inv_n - mean_diff_sq);
+}
+
+NK_PUBLIC void nk_kabsch_f16_haswell(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+                                     nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f32_t *result) {
+    // Fused single-pass: load f16, convert to f32, compute centroids and covariance
+    __m256 const zeros_f32x8 = _mm256_setzero_ps();
+
+    // Accumulators for centroids (f32)
+    __m256 sum_a_x_f32x8 = zeros_f32x8, sum_a_y_f32x8 = zeros_f32x8, sum_a_z_f32x8 = zeros_f32x8;
+    __m256 sum_b_x_f32x8 = zeros_f32x8, sum_b_y_f32x8 = zeros_f32x8, sum_b_z_f32x8 = zeros_f32x8;
+
+    // Accumulators for covariance matrix (sum of outer products)
+    __m256 cov_xx_f32x8 = zeros_f32x8, cov_xy_f32x8 = zeros_f32x8, cov_xz_f32x8 = zeros_f32x8;
+    __m256 cov_yx_f32x8 = zeros_f32x8, cov_yy_f32x8 = zeros_f32x8, cov_yz_f32x8 = zeros_f32x8;
+    __m256 cov_zx_f32x8 = zeros_f32x8, cov_zy_f32x8 = zeros_f32x8, cov_zz_f32x8 = zeros_f32x8;
+
+    nk_size_t i = 0;
+    __m256 a_x_f32x8, a_y_f32x8, a_z_f32x8, b_x_f32x8, b_y_f32x8, b_z_f32x8;
+
+    for (; i + 8 <= n; i += 8) {
+        nk_deinterleave_f16x8_to_f32x8_haswell_(a + i * 3, &a_x_f32x8, &a_y_f32x8, &a_z_f32x8);
+        nk_deinterleave_f16x8_to_f32x8_haswell_(b + i * 3, &b_x_f32x8, &b_y_f32x8, &b_z_f32x8);
+
+        // Accumulate centroids
+        sum_a_x_f32x8 = _mm256_add_ps(sum_a_x_f32x8, a_x_f32x8);
+        sum_a_y_f32x8 = _mm256_add_ps(sum_a_y_f32x8, a_y_f32x8);
+        sum_a_z_f32x8 = _mm256_add_ps(sum_a_z_f32x8, a_z_f32x8);
+        sum_b_x_f32x8 = _mm256_add_ps(sum_b_x_f32x8, b_x_f32x8);
+        sum_b_y_f32x8 = _mm256_add_ps(sum_b_y_f32x8, b_y_f32x8);
+        sum_b_z_f32x8 = _mm256_add_ps(sum_b_z_f32x8, b_z_f32x8);
+
+        // Accumulate outer products
+        cov_xx_f32x8 = _mm256_fmadd_ps(a_x_f32x8, b_x_f32x8, cov_xx_f32x8);
+        cov_xy_f32x8 = _mm256_fmadd_ps(a_x_f32x8, b_y_f32x8, cov_xy_f32x8);
+        cov_xz_f32x8 = _mm256_fmadd_ps(a_x_f32x8, b_z_f32x8, cov_xz_f32x8);
+        cov_yx_f32x8 = _mm256_fmadd_ps(a_y_f32x8, b_x_f32x8, cov_yx_f32x8);
+        cov_yy_f32x8 = _mm256_fmadd_ps(a_y_f32x8, b_y_f32x8, cov_yy_f32x8);
+        cov_yz_f32x8 = _mm256_fmadd_ps(a_y_f32x8, b_z_f32x8, cov_yz_f32x8);
+        cov_zx_f32x8 = _mm256_fmadd_ps(a_z_f32x8, b_x_f32x8, cov_zx_f32x8);
+        cov_zy_f32x8 = _mm256_fmadd_ps(a_z_f32x8, b_y_f32x8, cov_zy_f32x8);
+        cov_zz_f32x8 = _mm256_fmadd_ps(a_z_f32x8, b_z_f32x8, cov_zz_f32x8);
+    }
+
+    // Reduce vector accumulators
+    nk_f32_t sum_a_x = nk_reduce_add_f32x8_haswell_(sum_a_x_f32x8);
+    nk_f32_t sum_a_y = nk_reduce_add_f32x8_haswell_(sum_a_y_f32x8);
+    nk_f32_t sum_a_z = nk_reduce_add_f32x8_haswell_(sum_a_z_f32x8);
+    nk_f32_t sum_b_x = nk_reduce_add_f32x8_haswell_(sum_b_x_f32x8);
+    nk_f32_t sum_b_y = nk_reduce_add_f32x8_haswell_(sum_b_y_f32x8);
+    nk_f32_t sum_b_z = nk_reduce_add_f32x8_haswell_(sum_b_z_f32x8);
+
+    nk_f32_t covariance_x_x = nk_reduce_add_f32x8_haswell_(cov_xx_f32x8);
+    nk_f32_t covariance_x_y = nk_reduce_add_f32x8_haswell_(cov_xy_f32x8);
+    nk_f32_t covariance_x_z = nk_reduce_add_f32x8_haswell_(cov_xz_f32x8);
+    nk_f32_t covariance_y_x = nk_reduce_add_f32x8_haswell_(cov_yx_f32x8);
+    nk_f32_t covariance_y_y = nk_reduce_add_f32x8_haswell_(cov_yy_f32x8);
+    nk_f32_t covariance_y_z = nk_reduce_add_f32x8_haswell_(cov_yz_f32x8);
+    nk_f32_t covariance_z_x = nk_reduce_add_f32x8_haswell_(cov_zx_f32x8);
+    nk_f32_t covariance_z_y = nk_reduce_add_f32x8_haswell_(cov_zy_f32x8);
+    nk_f32_t covariance_z_z = nk_reduce_add_f32x8_haswell_(cov_zz_f32x8);
+
+    // Scalar tail
+    for (; i < n; ++i) {
+        nk_f32_t ax, ay, az, bx, by, bz;
+        nk_f16_to_f32_haswell(&a[i * 3 + 0], &ax);
+        nk_f16_to_f32_haswell(&a[i * 3 + 1], &ay);
+        nk_f16_to_f32_haswell(&a[i * 3 + 2], &az);
+        nk_f16_to_f32_haswell(&b[i * 3 + 0], &bx);
+        nk_f16_to_f32_haswell(&b[i * 3 + 1], &by);
+        nk_f16_to_f32_haswell(&b[i * 3 + 2], &bz);
+        sum_a_x += ax, sum_a_y += ay, sum_a_z += az;
+        sum_b_x += bx, sum_b_y += by, sum_b_z += bz;
+        covariance_x_x += ax * bx, covariance_x_y += ax * by, covariance_x_z += ax * bz;
+        covariance_y_x += ay * bx, covariance_y_y += ay * by, covariance_y_z += ay * bz;
+        covariance_z_x += az * bx, covariance_z_y += az * by, covariance_z_z += az * bz;
+    }
+
+    // Compute centroids
+    nk_f32_t inv_n = 1.0f / (nk_f32_t)n;
+    nk_f32_t centroid_a_x = sum_a_x * inv_n;
+    nk_f32_t centroid_a_y = sum_a_y * inv_n;
+    nk_f32_t centroid_a_z = sum_a_z * inv_n;
+    nk_f32_t centroid_b_x = sum_b_x * inv_n;
+    nk_f32_t centroid_b_y = sum_b_y * inv_n;
+    nk_f32_t centroid_b_z = sum_b_z * inv_n;
+
+    if (a_centroid) a_centroid[0] = centroid_a_x, a_centroid[1] = centroid_a_y, a_centroid[2] = centroid_a_z;
+    if (b_centroid) b_centroid[0] = centroid_b_x, b_centroid[1] = centroid_b_y, b_centroid[2] = centroid_b_z;
+
+    // Apply centering correction: H_centered = H - n * centroid_a * centroid_bᵀ
+    covariance_x_x -= (nk_f32_t)n * centroid_a_x * centroid_b_x;
+    covariance_x_y -= (nk_f32_t)n * centroid_a_x * centroid_b_y;
+    covariance_x_z -= (nk_f32_t)n * centroid_a_x * centroid_b_z;
+    covariance_y_x -= (nk_f32_t)n * centroid_a_y * centroid_b_x;
+    covariance_y_y -= (nk_f32_t)n * centroid_a_y * centroid_b_y;
+    covariance_y_z -= (nk_f32_t)n * centroid_a_y * centroid_b_z;
+    covariance_z_x -= (nk_f32_t)n * centroid_a_z * centroid_b_x;
+    covariance_z_y -= (nk_f32_t)n * centroid_a_z * centroid_b_y;
+    covariance_z_z -= (nk_f32_t)n * centroid_a_z * centroid_b_z;
+
+    // Compute SVD and optimal rotation
+    nk_f32_t cross_covariance[9] = {covariance_x_x, covariance_x_y, covariance_x_z, covariance_y_x, covariance_y_y,
+                                    covariance_y_z, covariance_z_x, covariance_z_y, covariance_z_z};
+    nk_f32_t svd_u[9], svd_s[9], svd_v[9];
+    nk_svd3x3_f32_(cross_covariance, svd_u, svd_s, svd_v);
+
+    // R = V * Uᵀ
+    nk_f32_t r[9];
+    r[0] = svd_v[0] * svd_u[0] + svd_v[1] * svd_u[1] + svd_v[2] * svd_u[2];
+    r[1] = svd_v[0] * svd_u[3] + svd_v[1] * svd_u[4] + svd_v[2] * svd_u[5];
+    r[2] = svd_v[0] * svd_u[6] + svd_v[1] * svd_u[7] + svd_v[2] * svd_u[8];
+    r[3] = svd_v[3] * svd_u[0] + svd_v[4] * svd_u[1] + svd_v[5] * svd_u[2];
+    r[4] = svd_v[3] * svd_u[3] + svd_v[4] * svd_u[4] + svd_v[5] * svd_u[5];
+    r[5] = svd_v[3] * svd_u[6] + svd_v[4] * svd_u[7] + svd_v[5] * svd_u[8];
+    r[6] = svd_v[6] * svd_u[0] + svd_v[7] * svd_u[1] + svd_v[8] * svd_u[2];
+    r[7] = svd_v[6] * svd_u[3] + svd_v[7] * svd_u[4] + svd_v[8] * svd_u[5];
+    r[8] = svd_v[6] * svd_u[6] + svd_v[7] * svd_u[7] + svd_v[8] * svd_u[8];
+
+    // Handle reflection: if det(R) < 0, negate third column of V and recompute R
+    if (nk_det3x3_f32_(r) < 0) {
+        svd_v[2] = -svd_v[2], svd_v[5] = -svd_v[5], svd_v[8] = -svd_v[8];
+        r[0] = svd_v[0] * svd_u[0] + svd_v[1] * svd_u[1] + svd_v[2] * svd_u[2];
+        r[1] = svd_v[0] * svd_u[3] + svd_v[1] * svd_u[4] + svd_v[2] * svd_u[5];
+        r[2] = svd_v[0] * svd_u[6] + svd_v[1] * svd_u[7] + svd_v[2] * svd_u[8];
+        r[3] = svd_v[3] * svd_u[0] + svd_v[4] * svd_u[1] + svd_v[5] * svd_u[2];
+        r[4] = svd_v[3] * svd_u[3] + svd_v[4] * svd_u[4] + svd_v[5] * svd_u[5];
+        r[5] = svd_v[3] * svd_u[6] + svd_v[4] * svd_u[7] + svd_v[5] * svd_u[8];
+        r[6] = svd_v[6] * svd_u[0] + svd_v[7] * svd_u[1] + svd_v[8] * svd_u[2];
+        r[7] = svd_v[6] * svd_u[3] + svd_v[7] * svd_u[4] + svd_v[8] * svd_u[5];
+        r[8] = svd_v[6] * svd_u[6] + svd_v[7] * svd_u[7] + svd_v[8] * svd_u[8];
+    }
+
+    // Output rotation matrix and scale=1.0
+    if (rotation)
+        for (int j = 0; j < 9; ++j) rotation[j] = r[j];
+    if (scale) *scale = 1.0f;
+
+    // Compute RMSD after optimal rotation
+    nk_f32_t sum_squared = nk_transformed_ssd_f16_haswell_(a, b, n, r, 1.0f, centroid_a_x, centroid_a_y, centroid_a_z,
+                                                           centroid_b_x, centroid_b_y, centroid_b_z);
+    *result = nk_f32_sqrt_haswell(sum_squared * inv_n);
+}
+
+NK_PUBLIC void nk_kabsch_bf16_haswell(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+                                      nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f32_t *result) {
+    // Fused single-pass: load bf16, convert to f32, compute centroids and covariance
+    __m256 const zeros_f32x8 = _mm256_setzero_ps();
+
+    // Accumulators for centroids (f32)
+    __m256 sum_a_x_f32x8 = zeros_f32x8, sum_a_y_f32x8 = zeros_f32x8, sum_a_z_f32x8 = zeros_f32x8;
+    __m256 sum_b_x_f32x8 = zeros_f32x8, sum_b_y_f32x8 = zeros_f32x8, sum_b_z_f32x8 = zeros_f32x8;
+
+    // Accumulators for covariance matrix (sum of outer products)
+    __m256 cov_xx_f32x8 = zeros_f32x8, cov_xy_f32x8 = zeros_f32x8, cov_xz_f32x8 = zeros_f32x8;
+    __m256 cov_yx_f32x8 = zeros_f32x8, cov_yy_f32x8 = zeros_f32x8, cov_yz_f32x8 = zeros_f32x8;
+    __m256 cov_zx_f32x8 = zeros_f32x8, cov_zy_f32x8 = zeros_f32x8, cov_zz_f32x8 = zeros_f32x8;
+
+    nk_size_t i = 0;
+    __m256 a_x_f32x8, a_y_f32x8, a_z_f32x8, b_x_f32x8, b_y_f32x8, b_z_f32x8;
+
+    for (; i + 8 <= n; i += 8) {
+        nk_deinterleave_bf16x8_to_f32x8_haswell_(a + i * 3, &a_x_f32x8, &a_y_f32x8, &a_z_f32x8);
+        nk_deinterleave_bf16x8_to_f32x8_haswell_(b + i * 3, &b_x_f32x8, &b_y_f32x8, &b_z_f32x8);
+
+        // Accumulate centroids
+        sum_a_x_f32x8 = _mm256_add_ps(sum_a_x_f32x8, a_x_f32x8);
+        sum_a_y_f32x8 = _mm256_add_ps(sum_a_y_f32x8, a_y_f32x8);
+        sum_a_z_f32x8 = _mm256_add_ps(sum_a_z_f32x8, a_z_f32x8);
+        sum_b_x_f32x8 = _mm256_add_ps(sum_b_x_f32x8, b_x_f32x8);
+        sum_b_y_f32x8 = _mm256_add_ps(sum_b_y_f32x8, b_y_f32x8);
+        sum_b_z_f32x8 = _mm256_add_ps(sum_b_z_f32x8, b_z_f32x8);
+
+        // Accumulate outer products
+        cov_xx_f32x8 = _mm256_fmadd_ps(a_x_f32x8, b_x_f32x8, cov_xx_f32x8);
+        cov_xy_f32x8 = _mm256_fmadd_ps(a_x_f32x8, b_y_f32x8, cov_xy_f32x8);
+        cov_xz_f32x8 = _mm256_fmadd_ps(a_x_f32x8, b_z_f32x8, cov_xz_f32x8);
+        cov_yx_f32x8 = _mm256_fmadd_ps(a_y_f32x8, b_x_f32x8, cov_yx_f32x8);
+        cov_yy_f32x8 = _mm256_fmadd_ps(a_y_f32x8, b_y_f32x8, cov_yy_f32x8);
+        cov_yz_f32x8 = _mm256_fmadd_ps(a_y_f32x8, b_z_f32x8, cov_yz_f32x8);
+        cov_zx_f32x8 = _mm256_fmadd_ps(a_z_f32x8, b_x_f32x8, cov_zx_f32x8);
+        cov_zy_f32x8 = _mm256_fmadd_ps(a_z_f32x8, b_y_f32x8, cov_zy_f32x8);
+        cov_zz_f32x8 = _mm256_fmadd_ps(a_z_f32x8, b_z_f32x8, cov_zz_f32x8);
+    }
+
+    // Reduce vector accumulators
+    nk_f32_t sum_a_x = nk_reduce_add_f32x8_haswell_(sum_a_x_f32x8);
+    nk_f32_t sum_a_y = nk_reduce_add_f32x8_haswell_(sum_a_y_f32x8);
+    nk_f32_t sum_a_z = nk_reduce_add_f32x8_haswell_(sum_a_z_f32x8);
+    nk_f32_t sum_b_x = nk_reduce_add_f32x8_haswell_(sum_b_x_f32x8);
+    nk_f32_t sum_b_y = nk_reduce_add_f32x8_haswell_(sum_b_y_f32x8);
+    nk_f32_t sum_b_z = nk_reduce_add_f32x8_haswell_(sum_b_z_f32x8);
+
+    nk_f32_t covariance_x_x = nk_reduce_add_f32x8_haswell_(cov_xx_f32x8);
+    nk_f32_t covariance_x_y = nk_reduce_add_f32x8_haswell_(cov_xy_f32x8);
+    nk_f32_t covariance_x_z = nk_reduce_add_f32x8_haswell_(cov_xz_f32x8);
+    nk_f32_t covariance_y_x = nk_reduce_add_f32x8_haswell_(cov_yx_f32x8);
+    nk_f32_t covariance_y_y = nk_reduce_add_f32x8_haswell_(cov_yy_f32x8);
+    nk_f32_t covariance_y_z = nk_reduce_add_f32x8_haswell_(cov_yz_f32x8);
+    nk_f32_t covariance_z_x = nk_reduce_add_f32x8_haswell_(cov_zx_f32x8);
+    nk_f32_t covariance_z_y = nk_reduce_add_f32x8_haswell_(cov_zy_f32x8);
+    nk_f32_t covariance_z_z = nk_reduce_add_f32x8_haswell_(cov_zz_f32x8);
+
+    // Scalar tail
+    for (; i < n; ++i) {
+        nk_f32_t ax, ay, az, bx, by, bz;
+        nk_bf16_to_f32_serial(&a[i * 3 + 0], &ax);
+        nk_bf16_to_f32_serial(&a[i * 3 + 1], &ay);
+        nk_bf16_to_f32_serial(&a[i * 3 + 2], &az);
+        nk_bf16_to_f32_serial(&b[i * 3 + 0], &bx);
+        nk_bf16_to_f32_serial(&b[i * 3 + 1], &by);
+        nk_bf16_to_f32_serial(&b[i * 3 + 2], &bz);
+        sum_a_x += ax, sum_a_y += ay, sum_a_z += az;
+        sum_b_x += bx, sum_b_y += by, sum_b_z += bz;
+        covariance_x_x += ax * bx, covariance_x_y += ax * by, covariance_x_z += ax * bz;
+        covariance_y_x += ay * bx, covariance_y_y += ay * by, covariance_y_z += ay * bz;
+        covariance_z_x += az * bx, covariance_z_y += az * by, covariance_z_z += az * bz;
+    }
+
+    // Compute centroids
+    nk_f32_t inv_n = 1.0f / (nk_f32_t)n;
+    nk_f32_t centroid_a_x = sum_a_x * inv_n;
+    nk_f32_t centroid_a_y = sum_a_y * inv_n;
+    nk_f32_t centroid_a_z = sum_a_z * inv_n;
+    nk_f32_t centroid_b_x = sum_b_x * inv_n;
+    nk_f32_t centroid_b_y = sum_b_y * inv_n;
+    nk_f32_t centroid_b_z = sum_b_z * inv_n;
+
+    if (a_centroid) a_centroid[0] = centroid_a_x, a_centroid[1] = centroid_a_y, a_centroid[2] = centroid_a_z;
+    if (b_centroid) b_centroid[0] = centroid_b_x, b_centroid[1] = centroid_b_y, b_centroid[2] = centroid_b_z;
+
+    // Apply centering correction: H_centered = H - n * centroid_a * centroid_bᵀ
+    covariance_x_x -= (nk_f32_t)n * centroid_a_x * centroid_b_x;
+    covariance_x_y -= (nk_f32_t)n * centroid_a_x * centroid_b_y;
+    covariance_x_z -= (nk_f32_t)n * centroid_a_x * centroid_b_z;
+    covariance_y_x -= (nk_f32_t)n * centroid_a_y * centroid_b_x;
+    covariance_y_y -= (nk_f32_t)n * centroid_a_y * centroid_b_y;
+    covariance_y_z -= (nk_f32_t)n * centroid_a_y * centroid_b_z;
+    covariance_z_x -= (nk_f32_t)n * centroid_a_z * centroid_b_x;
+    covariance_z_y -= (nk_f32_t)n * centroid_a_z * centroid_b_y;
+    covariance_z_z -= (nk_f32_t)n * centroid_a_z * centroid_b_z;
+
+    // Compute SVD and optimal rotation
+    nk_f32_t cross_covariance[9] = {covariance_x_x, covariance_x_y, covariance_x_z, covariance_y_x, covariance_y_y,
+                                    covariance_y_z, covariance_z_x, covariance_z_y, covariance_z_z};
+    nk_f32_t svd_u[9], svd_s[9], svd_v[9];
+    nk_svd3x3_f32_(cross_covariance, svd_u, svd_s, svd_v);
+
+    // R = V * Uᵀ
+    nk_f32_t r[9];
+    r[0] = svd_v[0] * svd_u[0] + svd_v[1] * svd_u[1] + svd_v[2] * svd_u[2];
+    r[1] = svd_v[0] * svd_u[3] + svd_v[1] * svd_u[4] + svd_v[2] * svd_u[5];
+    r[2] = svd_v[0] * svd_u[6] + svd_v[1] * svd_u[7] + svd_v[2] * svd_u[8];
+    r[3] = svd_v[3] * svd_u[0] + svd_v[4] * svd_u[1] + svd_v[5] * svd_u[2];
+    r[4] = svd_v[3] * svd_u[3] + svd_v[4] * svd_u[4] + svd_v[5] * svd_u[5];
+    r[5] = svd_v[3] * svd_u[6] + svd_v[4] * svd_u[7] + svd_v[5] * svd_u[8];
+    r[6] = svd_v[6] * svd_u[0] + svd_v[7] * svd_u[1] + svd_v[8] * svd_u[2];
+    r[7] = svd_v[6] * svd_u[3] + svd_v[7] * svd_u[4] + svd_v[8] * svd_u[5];
+    r[8] = svd_v[6] * svd_u[6] + svd_v[7] * svd_u[7] + svd_v[8] * svd_u[8];
+
+    // Handle reflection: if det(R) < 0, negate third column of V and recompute R
+    if (nk_det3x3_f32_(r) < 0) {
+        svd_v[2] = -svd_v[2], svd_v[5] = -svd_v[5], svd_v[8] = -svd_v[8];
+        r[0] = svd_v[0] * svd_u[0] + svd_v[1] * svd_u[1] + svd_v[2] * svd_u[2];
+        r[1] = svd_v[0] * svd_u[3] + svd_v[1] * svd_u[4] + svd_v[2] * svd_u[5];
+        r[2] = svd_v[0] * svd_u[6] + svd_v[1] * svd_u[7] + svd_v[2] * svd_u[8];
+        r[3] = svd_v[3] * svd_u[0] + svd_v[4] * svd_u[1] + svd_v[5] * svd_u[2];
+        r[4] = svd_v[3] * svd_u[3] + svd_v[4] * svd_u[4] + svd_v[5] * svd_u[5];
+        r[5] = svd_v[3] * svd_u[6] + svd_v[4] * svd_u[7] + svd_v[5] * svd_u[8];
+        r[6] = svd_v[6] * svd_u[0] + svd_v[7] * svd_u[1] + svd_v[8] * svd_u[2];
+        r[7] = svd_v[6] * svd_u[3] + svd_v[7] * svd_u[4] + svd_v[8] * svd_u[5];
+        r[8] = svd_v[6] * svd_u[6] + svd_v[7] * svd_u[7] + svd_v[8] * svd_u[8];
+    }
+
+    // Output rotation matrix and scale=1.0
+    if (rotation)
+        for (int j = 0; j < 9; ++j) rotation[j] = r[j];
+    if (scale) *scale = 1.0f;
+
+    // Compute RMSD after optimal rotation
+    nk_f32_t sum_squared = nk_transformed_ssd_bf16_haswell_(a, b, n, r, 1.0f, centroid_a_x, centroid_a_y, centroid_a_z,
+                                                            centroid_b_x, centroid_b_y, centroid_b_z);
+    *result = nk_f32_sqrt_haswell(sum_squared * inv_n);
+}
+
+NK_PUBLIC void nk_umeyama_f16_haswell(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+                                      nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f32_t *result) {
+    // Fused single-pass: load f16, convert to f32, compute centroids, covariance, and variance
+    __m256 const zeros_f32x8 = _mm256_setzero_ps();
+
+    __m256 sum_a_x_f32x8 = zeros_f32x8, sum_a_y_f32x8 = zeros_f32x8, sum_a_z_f32x8 = zeros_f32x8;
+    __m256 sum_b_x_f32x8 = zeros_f32x8, sum_b_y_f32x8 = zeros_f32x8, sum_b_z_f32x8 = zeros_f32x8;
+    __m256 cov_xx_f32x8 = zeros_f32x8, cov_xy_f32x8 = zeros_f32x8, cov_xz_f32x8 = zeros_f32x8;
+    __m256 cov_yx_f32x8 = zeros_f32x8, cov_yy_f32x8 = zeros_f32x8, cov_yz_f32x8 = zeros_f32x8;
+    __m256 cov_zx_f32x8 = zeros_f32x8, cov_zy_f32x8 = zeros_f32x8, cov_zz_f32x8 = zeros_f32x8;
+    __m256 variance_a_f32x8 = zeros_f32x8;
+
+    nk_size_t i = 0;
+    __m256 a_x_f32x8, a_y_f32x8, a_z_f32x8, b_x_f32x8, b_y_f32x8, b_z_f32x8;
+
+    for (; i + 8 <= n; i += 8) {
+        nk_deinterleave_f16x8_to_f32x8_haswell_(a + i * 3, &a_x_f32x8, &a_y_f32x8, &a_z_f32x8);
+        nk_deinterleave_f16x8_to_f32x8_haswell_(b + i * 3, &b_x_f32x8, &b_y_f32x8, &b_z_f32x8);
+
+        // Accumulate centroids
+        sum_a_x_f32x8 = _mm256_add_ps(sum_a_x_f32x8, a_x_f32x8);
+        sum_a_y_f32x8 = _mm256_add_ps(sum_a_y_f32x8, a_y_f32x8);
+        sum_a_z_f32x8 = _mm256_add_ps(sum_a_z_f32x8, a_z_f32x8);
+        sum_b_x_f32x8 = _mm256_add_ps(sum_b_x_f32x8, b_x_f32x8);
+        sum_b_y_f32x8 = _mm256_add_ps(sum_b_y_f32x8, b_y_f32x8);
+        sum_b_z_f32x8 = _mm256_add_ps(sum_b_z_f32x8, b_z_f32x8);
+
+        // Accumulate outer products
+        cov_xx_f32x8 = _mm256_fmadd_ps(a_x_f32x8, b_x_f32x8, cov_xx_f32x8);
+        cov_xy_f32x8 = _mm256_fmadd_ps(a_x_f32x8, b_y_f32x8, cov_xy_f32x8);
+        cov_xz_f32x8 = _mm256_fmadd_ps(a_x_f32x8, b_z_f32x8, cov_xz_f32x8);
+        cov_yx_f32x8 = _mm256_fmadd_ps(a_y_f32x8, b_x_f32x8, cov_yx_f32x8);
+        cov_yy_f32x8 = _mm256_fmadd_ps(a_y_f32x8, b_y_f32x8, cov_yy_f32x8);
+        cov_yz_f32x8 = _mm256_fmadd_ps(a_y_f32x8, b_z_f32x8, cov_yz_f32x8);
+        cov_zx_f32x8 = _mm256_fmadd_ps(a_z_f32x8, b_x_f32x8, cov_zx_f32x8);
+        cov_zy_f32x8 = _mm256_fmadd_ps(a_z_f32x8, b_y_f32x8, cov_zy_f32x8);
+        cov_zz_f32x8 = _mm256_fmadd_ps(a_z_f32x8, b_z_f32x8, cov_zz_f32x8);
+
+        // Accumulate variance of A
+        variance_a_f32x8 = _mm256_fmadd_ps(a_x_f32x8, a_x_f32x8, variance_a_f32x8);
+        variance_a_f32x8 = _mm256_fmadd_ps(a_y_f32x8, a_y_f32x8, variance_a_f32x8);
+        variance_a_f32x8 = _mm256_fmadd_ps(a_z_f32x8, a_z_f32x8, variance_a_f32x8);
+    }
+
+    // Reduce vector accumulators
+    nk_f32_t sum_a_x = nk_reduce_add_f32x8_haswell_(sum_a_x_f32x8);
+    nk_f32_t sum_a_y = nk_reduce_add_f32x8_haswell_(sum_a_y_f32x8);
+    nk_f32_t sum_a_z = nk_reduce_add_f32x8_haswell_(sum_a_z_f32x8);
+    nk_f32_t sum_b_x = nk_reduce_add_f32x8_haswell_(sum_b_x_f32x8);
+    nk_f32_t sum_b_y = nk_reduce_add_f32x8_haswell_(sum_b_y_f32x8);
+    nk_f32_t sum_b_z = nk_reduce_add_f32x8_haswell_(sum_b_z_f32x8);
+    nk_f32_t covariance_x_x = nk_reduce_add_f32x8_haswell_(cov_xx_f32x8);
+    nk_f32_t covariance_x_y = nk_reduce_add_f32x8_haswell_(cov_xy_f32x8);
+    nk_f32_t covariance_x_z = nk_reduce_add_f32x8_haswell_(cov_xz_f32x8);
+    nk_f32_t covariance_y_x = nk_reduce_add_f32x8_haswell_(cov_yx_f32x8);
+    nk_f32_t covariance_y_y = nk_reduce_add_f32x8_haswell_(cov_yy_f32x8);
+    nk_f32_t covariance_y_z = nk_reduce_add_f32x8_haswell_(cov_yz_f32x8);
+    nk_f32_t covariance_z_x = nk_reduce_add_f32x8_haswell_(cov_zx_f32x8);
+    nk_f32_t covariance_z_y = nk_reduce_add_f32x8_haswell_(cov_zy_f32x8);
+    nk_f32_t covariance_z_z = nk_reduce_add_f32x8_haswell_(cov_zz_f32x8);
+    nk_f32_t variance_a_sum = nk_reduce_add_f32x8_haswell_(variance_a_f32x8);
+
+    // Scalar tail
+    for (; i < n; ++i) {
+        nk_f32_t ax, ay, az, bx, by, bz;
+        nk_f16_to_f32_haswell(&a[i * 3 + 0], &ax);
+        nk_f16_to_f32_haswell(&a[i * 3 + 1], &ay);
+        nk_f16_to_f32_haswell(&a[i * 3 + 2], &az);
+        nk_f16_to_f32_haswell(&b[i * 3 + 0], &bx);
+        nk_f16_to_f32_haswell(&b[i * 3 + 1], &by);
+        nk_f16_to_f32_haswell(&b[i * 3 + 2], &bz);
+        sum_a_x += ax, sum_a_y += ay, sum_a_z += az;
+        sum_b_x += bx, sum_b_y += by, sum_b_z += bz;
+        covariance_x_x += ax * bx, covariance_x_y += ax * by, covariance_x_z += ax * bz;
+        covariance_y_x += ay * bx, covariance_y_y += ay * by, covariance_y_z += ay * bz;
+        covariance_z_x += az * bx, covariance_z_y += az * by, covariance_z_z += az * bz;
+        variance_a_sum += ax * ax + ay * ay + az * az;
+    }
+
+    // Compute centroids
+    nk_f32_t inv_n = 1.0f / (nk_f32_t)n;
+    nk_f32_t centroid_a_x = sum_a_x * inv_n, centroid_a_y = sum_a_y * inv_n, centroid_a_z = sum_a_z * inv_n;
+    nk_f32_t centroid_b_x = sum_b_x * inv_n, centroid_b_y = sum_b_y * inv_n, centroid_b_z = sum_b_z * inv_n;
+
+    if (a_centroid) a_centroid[0] = centroid_a_x, a_centroid[1] = centroid_a_y, a_centroid[2] = centroid_a_z;
+    if (b_centroid) b_centroid[0] = centroid_b_x, b_centroid[1] = centroid_b_y, b_centroid[2] = centroid_b_z;
+
+    // Compute centered covariance and variance
+    nk_f32_t variance_a = variance_a_sum * inv_n -
+                          (centroid_a_x * centroid_a_x + centroid_a_y * centroid_a_y + centroid_a_z * centroid_a_z);
+
+    // Apply centering correction to covariance matrix
+    covariance_x_x -= (nk_f32_t)n * centroid_a_x * centroid_b_x;
+    covariance_x_y -= (nk_f32_t)n * centroid_a_x * centroid_b_y;
+    covariance_x_z -= (nk_f32_t)n * centroid_a_x * centroid_b_z;
+    covariance_y_x -= (nk_f32_t)n * centroid_a_y * centroid_b_x;
+    covariance_y_y -= (nk_f32_t)n * centroid_a_y * centroid_b_y;
+    covariance_y_z -= (nk_f32_t)n * centroid_a_y * centroid_b_z;
+    covariance_z_x -= (nk_f32_t)n * centroid_a_z * centroid_b_x;
+    covariance_z_y -= (nk_f32_t)n * centroid_a_z * centroid_b_y;
+    covariance_z_z -= (nk_f32_t)n * centroid_a_z * centroid_b_z;
+
+    nk_f32_t cross_covariance[9] = {covariance_x_x, covariance_x_y, covariance_x_z, covariance_y_x, covariance_y_y,
+                                    covariance_y_z, covariance_z_x, covariance_z_y, covariance_z_z};
+
+    // SVD
+    nk_f32_t svd_u[9], svd_s[9], svd_v[9];
+    nk_svd3x3_f32_(cross_covariance, svd_u, svd_s, svd_v);
+
+    // R = V * Uᵀ
+    nk_f32_t r[9];
+    r[0] = svd_v[0] * svd_u[0] + svd_v[1] * svd_u[1] + svd_v[2] * svd_u[2];
+    r[1] = svd_v[0] * svd_u[3] + svd_v[1] * svd_u[4] + svd_v[2] * svd_u[5];
+    r[2] = svd_v[0] * svd_u[6] + svd_v[1] * svd_u[7] + svd_v[2] * svd_u[8];
+    r[3] = svd_v[3] * svd_u[0] + svd_v[4] * svd_u[1] + svd_v[5] * svd_u[2];
+    r[4] = svd_v[3] * svd_u[3] + svd_v[4] * svd_u[4] + svd_v[5] * svd_u[5];
+    r[5] = svd_v[3] * svd_u[6] + svd_v[4] * svd_u[7] + svd_v[5] * svd_u[8];
+    r[6] = svd_v[6] * svd_u[0] + svd_v[7] * svd_u[1] + svd_v[8] * svd_u[2];
+    r[7] = svd_v[6] * svd_u[3] + svd_v[7] * svd_u[4] + svd_v[8] * svd_u[5];
+    r[8] = svd_v[6] * svd_u[6] + svd_v[7] * svd_u[7] + svd_v[8] * svd_u[8];
+
+    // Scale factor: c = trace(D × S) / (n × variance(a))
+    nk_f32_t det = nk_det3x3_f32_(r);
+    nk_f32_t d3 = det < 0 ? -1.0f : 1.0f;
+    nk_f32_t trace_ds = svd_s[0] + svd_s[4] + d3 * svd_s[8];
+    nk_f32_t c = trace_ds / ((nk_f32_t)n * variance_a);
+    if (scale) *scale = c;
+
+    // Handle reflection
+    if (det < 0) {
+        svd_v[2] = -svd_v[2], svd_v[5] = -svd_v[5], svd_v[8] = -svd_v[8];
+        r[0] = svd_v[0] * svd_u[0] + svd_v[1] * svd_u[1] + svd_v[2] * svd_u[2];
+        r[1] = svd_v[0] * svd_u[3] + svd_v[1] * svd_u[4] + svd_v[2] * svd_u[5];
+        r[2] = svd_v[0] * svd_u[6] + svd_v[1] * svd_u[7] + svd_v[2] * svd_u[8];
+        r[3] = svd_v[3] * svd_u[0] + svd_v[4] * svd_u[1] + svd_v[5] * svd_u[2];
+        r[4] = svd_v[3] * svd_u[3] + svd_v[4] * svd_u[4] + svd_v[5] * svd_u[5];
+        r[5] = svd_v[3] * svd_u[6] + svd_v[4] * svd_u[7] + svd_v[5] * svd_u[8];
+        r[6] = svd_v[6] * svd_u[0] + svd_v[7] * svd_u[1] + svd_v[8] * svd_u[2];
+        r[7] = svd_v[6] * svd_u[3] + svd_v[7] * svd_u[4] + svd_v[8] * svd_u[5];
+        r[8] = svd_v[6] * svd_u[6] + svd_v[7] * svd_u[7] + svd_v[8] * svd_u[8];
+    }
+
+    // Output rotation matrix
+    if (rotation)
+        for (int j = 0; j < 9; ++j) rotation[j] = r[j];
+
+    // Compute RMSD with scaling
+    nk_f32_t sum_squared = nk_transformed_ssd_f16_haswell_(a, b, n, r, c, centroid_a_x, centroid_a_y, centroid_a_z,
+                                                           centroid_b_x, centroid_b_y, centroid_b_z);
+    *result = nk_f32_sqrt_haswell(sum_squared * inv_n);
+}
+
+NK_PUBLIC void nk_umeyama_bf16_haswell(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+                                       nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f32_t *result) {
+    // Fused single-pass: load bf16, convert to f32, compute centroids, covariance, and variance
+    __m256 const zeros_f32x8 = _mm256_setzero_ps();
+
+    __m256 sum_a_x_f32x8 = zeros_f32x8, sum_a_y_f32x8 = zeros_f32x8, sum_a_z_f32x8 = zeros_f32x8;
+    __m256 sum_b_x_f32x8 = zeros_f32x8, sum_b_y_f32x8 = zeros_f32x8, sum_b_z_f32x8 = zeros_f32x8;
+    __m256 cov_xx_f32x8 = zeros_f32x8, cov_xy_f32x8 = zeros_f32x8, cov_xz_f32x8 = zeros_f32x8;
+    __m256 cov_yx_f32x8 = zeros_f32x8, cov_yy_f32x8 = zeros_f32x8, cov_yz_f32x8 = zeros_f32x8;
+    __m256 cov_zx_f32x8 = zeros_f32x8, cov_zy_f32x8 = zeros_f32x8, cov_zz_f32x8 = zeros_f32x8;
+    __m256 variance_a_f32x8 = zeros_f32x8;
+
+    nk_size_t i = 0;
+    __m256 a_x_f32x8, a_y_f32x8, a_z_f32x8, b_x_f32x8, b_y_f32x8, b_z_f32x8;
+
+    for (; i + 8 <= n; i += 8) {
+        nk_deinterleave_bf16x8_to_f32x8_haswell_(a + i * 3, &a_x_f32x8, &a_y_f32x8, &a_z_f32x8);
+        nk_deinterleave_bf16x8_to_f32x8_haswell_(b + i * 3, &b_x_f32x8, &b_y_f32x8, &b_z_f32x8);
+
+        // Accumulate centroids
+        sum_a_x_f32x8 = _mm256_add_ps(sum_a_x_f32x8, a_x_f32x8);
+        sum_a_y_f32x8 = _mm256_add_ps(sum_a_y_f32x8, a_y_f32x8);
+        sum_a_z_f32x8 = _mm256_add_ps(sum_a_z_f32x8, a_z_f32x8);
+        sum_b_x_f32x8 = _mm256_add_ps(sum_b_x_f32x8, b_x_f32x8);
+        sum_b_y_f32x8 = _mm256_add_ps(sum_b_y_f32x8, b_y_f32x8);
+        sum_b_z_f32x8 = _mm256_add_ps(sum_b_z_f32x8, b_z_f32x8);
+
+        // Accumulate outer products
+        cov_xx_f32x8 = _mm256_fmadd_ps(a_x_f32x8, b_x_f32x8, cov_xx_f32x8);
+        cov_xy_f32x8 = _mm256_fmadd_ps(a_x_f32x8, b_y_f32x8, cov_xy_f32x8);
+        cov_xz_f32x8 = _mm256_fmadd_ps(a_x_f32x8, b_z_f32x8, cov_xz_f32x8);
+        cov_yx_f32x8 = _mm256_fmadd_ps(a_y_f32x8, b_x_f32x8, cov_yx_f32x8);
+        cov_yy_f32x8 = _mm256_fmadd_ps(a_y_f32x8, b_y_f32x8, cov_yy_f32x8);
+        cov_yz_f32x8 = _mm256_fmadd_ps(a_y_f32x8, b_z_f32x8, cov_yz_f32x8);
+        cov_zx_f32x8 = _mm256_fmadd_ps(a_z_f32x8, b_x_f32x8, cov_zx_f32x8);
+        cov_zy_f32x8 = _mm256_fmadd_ps(a_z_f32x8, b_y_f32x8, cov_zy_f32x8);
+        cov_zz_f32x8 = _mm256_fmadd_ps(a_z_f32x8, b_z_f32x8, cov_zz_f32x8);
+
+        // Accumulate variance of A
+        variance_a_f32x8 = _mm256_fmadd_ps(a_x_f32x8, a_x_f32x8, variance_a_f32x8);
+        variance_a_f32x8 = _mm256_fmadd_ps(a_y_f32x8, a_y_f32x8, variance_a_f32x8);
+        variance_a_f32x8 = _mm256_fmadd_ps(a_z_f32x8, a_z_f32x8, variance_a_f32x8);
+    }
+
+    // Reduce vector accumulators
+    nk_f32_t sum_a_x = nk_reduce_add_f32x8_haswell_(sum_a_x_f32x8);
+    nk_f32_t sum_a_y = nk_reduce_add_f32x8_haswell_(sum_a_y_f32x8);
+    nk_f32_t sum_a_z = nk_reduce_add_f32x8_haswell_(sum_a_z_f32x8);
+    nk_f32_t sum_b_x = nk_reduce_add_f32x8_haswell_(sum_b_x_f32x8);
+    nk_f32_t sum_b_y = nk_reduce_add_f32x8_haswell_(sum_b_y_f32x8);
+    nk_f32_t sum_b_z = nk_reduce_add_f32x8_haswell_(sum_b_z_f32x8);
+    nk_f32_t covariance_x_x = nk_reduce_add_f32x8_haswell_(cov_xx_f32x8);
+    nk_f32_t covariance_x_y = nk_reduce_add_f32x8_haswell_(cov_xy_f32x8);
+    nk_f32_t covariance_x_z = nk_reduce_add_f32x8_haswell_(cov_xz_f32x8);
+    nk_f32_t covariance_y_x = nk_reduce_add_f32x8_haswell_(cov_yx_f32x8);
+    nk_f32_t covariance_y_y = nk_reduce_add_f32x8_haswell_(cov_yy_f32x8);
+    nk_f32_t covariance_y_z = nk_reduce_add_f32x8_haswell_(cov_yz_f32x8);
+    nk_f32_t covariance_z_x = nk_reduce_add_f32x8_haswell_(cov_zx_f32x8);
+    nk_f32_t covariance_z_y = nk_reduce_add_f32x8_haswell_(cov_zy_f32x8);
+    nk_f32_t covariance_z_z = nk_reduce_add_f32x8_haswell_(cov_zz_f32x8);
+    nk_f32_t variance_a_sum = nk_reduce_add_f32x8_haswell_(variance_a_f32x8);
+
+    // Scalar tail
+    for (; i < n; ++i) {
+        nk_f32_t ax, ay, az, bx, by, bz;
+        nk_bf16_to_f32_serial(&a[i * 3 + 0], &ax);
+        nk_bf16_to_f32_serial(&a[i * 3 + 1], &ay);
+        nk_bf16_to_f32_serial(&a[i * 3 + 2], &az);
+        nk_bf16_to_f32_serial(&b[i * 3 + 0], &bx);
+        nk_bf16_to_f32_serial(&b[i * 3 + 1], &by);
+        nk_bf16_to_f32_serial(&b[i * 3 + 2], &bz);
+        sum_a_x += ax, sum_a_y += ay, sum_a_z += az;
+        sum_b_x += bx, sum_b_y += by, sum_b_z += bz;
+        covariance_x_x += ax * bx, covariance_x_y += ax * by, covariance_x_z += ax * bz;
+        covariance_y_x += ay * bx, covariance_y_y += ay * by, covariance_y_z += ay * bz;
+        covariance_z_x += az * bx, covariance_z_y += az * by, covariance_z_z += az * bz;
+        variance_a_sum += ax * ax + ay * ay + az * az;
+    }
+
+    // Compute centroids
+    nk_f32_t inv_n = 1.0f / (nk_f32_t)n;
+    nk_f32_t centroid_a_x = sum_a_x * inv_n, centroid_a_y = sum_a_y * inv_n, centroid_a_z = sum_a_z * inv_n;
+    nk_f32_t centroid_b_x = sum_b_x * inv_n, centroid_b_y = sum_b_y * inv_n, centroid_b_z = sum_b_z * inv_n;
+
+    if (a_centroid) a_centroid[0] = centroid_a_x, a_centroid[1] = centroid_a_y, a_centroid[2] = centroid_a_z;
+    if (b_centroid) b_centroid[0] = centroid_b_x, b_centroid[1] = centroid_b_y, b_centroid[2] = centroid_b_z;
+
+    // Compute centered covariance and variance
+    nk_f32_t variance_a = variance_a_sum * inv_n -
+                          (centroid_a_x * centroid_a_x + centroid_a_y * centroid_a_y + centroid_a_z * centroid_a_z);
+
+    // Apply centering correction to covariance matrix
+    covariance_x_x -= (nk_f32_t)n * centroid_a_x * centroid_b_x;
+    covariance_x_y -= (nk_f32_t)n * centroid_a_x * centroid_b_y;
+    covariance_x_z -= (nk_f32_t)n * centroid_a_x * centroid_b_z;
+    covariance_y_x -= (nk_f32_t)n * centroid_a_y * centroid_b_x;
+    covariance_y_y -= (nk_f32_t)n * centroid_a_y * centroid_b_y;
+    covariance_y_z -= (nk_f32_t)n * centroid_a_y * centroid_b_z;
+    covariance_z_x -= (nk_f32_t)n * centroid_a_z * centroid_b_x;
+    covariance_z_y -= (nk_f32_t)n * centroid_a_z * centroid_b_y;
+    covariance_z_z -= (nk_f32_t)n * centroid_a_z * centroid_b_z;
+
+    nk_f32_t cross_covariance[9] = {covariance_x_x, covariance_x_y, covariance_x_z, covariance_y_x, covariance_y_y,
+                                    covariance_y_z, covariance_z_x, covariance_z_y, covariance_z_z};
+
+    // SVD
+    nk_f32_t svd_u[9], svd_s[9], svd_v[9];
+    nk_svd3x3_f32_(cross_covariance, svd_u, svd_s, svd_v);
+
+    // R = V * Uᵀ
+    nk_f32_t r[9];
+    r[0] = svd_v[0] * svd_u[0] + svd_v[1] * svd_u[1] + svd_v[2] * svd_u[2];
+    r[1] = svd_v[0] * svd_u[3] + svd_v[1] * svd_u[4] + svd_v[2] * svd_u[5];
+    r[2] = svd_v[0] * svd_u[6] + svd_v[1] * svd_u[7] + svd_v[2] * svd_u[8];
+    r[3] = svd_v[3] * svd_u[0] + svd_v[4] * svd_u[1] + svd_v[5] * svd_u[2];
+    r[4] = svd_v[3] * svd_u[3] + svd_v[4] * svd_u[4] + svd_v[5] * svd_u[5];
+    r[5] = svd_v[3] * svd_u[6] + svd_v[4] * svd_u[7] + svd_v[5] * svd_u[8];
+    r[6] = svd_v[6] * svd_u[0] + svd_v[7] * svd_u[1] + svd_v[8] * svd_u[2];
+    r[7] = svd_v[6] * svd_u[3] + svd_v[7] * svd_u[4] + svd_v[8] * svd_u[5];
+    r[8] = svd_v[6] * svd_u[6] + svd_v[7] * svd_u[7] + svd_v[8] * svd_u[8];
+
+    // Scale factor: c = trace(D × S) / (n × variance(a))
+    nk_f32_t det = nk_det3x3_f32_(r);
+    nk_f32_t d3 = det < 0 ? -1.0f : 1.0f;
+    nk_f32_t trace_ds = svd_s[0] + svd_s[4] + d3 * svd_s[8];
+    nk_f32_t c = trace_ds / ((nk_f32_t)n * variance_a);
+    if (scale) *scale = c;
+
+    // Handle reflection
+    if (det < 0) {
+        svd_v[2] = -svd_v[2], svd_v[5] = -svd_v[5], svd_v[8] = -svd_v[8];
+        r[0] = svd_v[0] * svd_u[0] + svd_v[1] * svd_u[1] + svd_v[2] * svd_u[2];
+        r[1] = svd_v[0] * svd_u[3] + svd_v[1] * svd_u[4] + svd_v[2] * svd_u[5];
+        r[2] = svd_v[0] * svd_u[6] + svd_v[1] * svd_u[7] + svd_v[2] * svd_u[8];
+        r[3] = svd_v[3] * svd_u[0] + svd_v[4] * svd_u[1] + svd_v[5] * svd_u[2];
+        r[4] = svd_v[3] * svd_u[3] + svd_v[4] * svd_u[4] + svd_v[5] * svd_u[5];
+        r[5] = svd_v[3] * svd_u[6] + svd_v[4] * svd_u[7] + svd_v[5] * svd_u[8];
+        r[6] = svd_v[6] * svd_u[0] + svd_v[7] * svd_u[1] + svd_v[8] * svd_u[2];
+        r[7] = svd_v[6] * svd_u[3] + svd_v[7] * svd_u[4] + svd_v[8] * svd_u[5];
+        r[8] = svd_v[6] * svd_u[6] + svd_v[7] * svd_u[7] + svd_v[8] * svd_u[8];
+    }
+
+    // Output rotation matrix
+    if (rotation)
+        for (int j = 0; j < 9; ++j) rotation[j] = r[j];
+
+    // Compute RMSD with scaling
+    nk_f32_t sum_squared = nk_transformed_ssd_bf16_haswell_(a, b, n, r, c, centroid_a_x, centroid_a_y, centroid_a_z,
+                                                            centroid_b_x, centroid_b_y, centroid_b_z);
+    *result = nk_f32_sqrt_haswell(sum_squared * inv_n);
+}
+
+#if defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#endif // NK_TARGET_HASWELL
+#endif // NK_TARGET_X8664_
+#endif // NK_MESH_HASWELL_H